Object Detection Models

Gemini 3.5 Flash-Lite is a natively multimodal reasoning model developed by Google DeepMind, released on July 21, 2026 as part of the Gemini 3.5 model family. It is the fastest model in the 3.5 series, designed for both low-latency tasks and high-throughput production workloads such as agentic search, document processing, receipt translation, and large-scale data extraction. The model accepts text, images, audio, and video as inputs, with a context window of up to 1 million tokens, and produces text output. It supports configurable thinking levels, allowing developers to tune the balance between response quality, cost, and latency depending on workload requirements.On agentic and coding benchmarks, Gemini 3.5 Flash-Lite significantly outperforms its predecessor, Gemini 3.1 Flash-Lite, including on Terminal-Bench 2.1 (54% vs. 31%), GDM-MRCR v2 long-context (72.2% vs. 60.1%), and real-world task execution as measured by GDPval-AA v2 (1140 vs. 642). It also surpasses Gemini 3 Flash on SWE-Bench Pro (54.2% vs. 49.6%) and OSWorld-Verified (74.0% vs. 65.1%). According to the Artificial Analysis Index, the model generates output at approximately 350 tokens per second. It is built on the Gemini 3.5 Flash foundation and is evaluated across reasoning, coding, multimodal understanding, multilingual performance, and long-context tasks. The model is developed under Google's Frontier Safety Framework.

Gemini 3.6 Flash is a multimodal language model from Google DeepMind, positioned as the workhorse tier in the Gemini 3.x family. It accepts text, image, video, audio, and PDF inputs with a 1 million token context window and produces up to 64,000 output tokens. The model builds directly on Gemini 3.5 Flash, incorporating developer and customer feedback to improve token efficiency, coding quality, and knowledge work performance. According to the Artificial Analysis Index, it consumes 17% fewer output tokens than its predecessor, and on some benchmarks such as DeepSWE, token reduction reaches up to 65%. It supports function calling, structured output, search as a tool, and code execution, and includes computer use as a built-in capability in the Gemini API and Gemini Enterprise.On coding benchmarks, Gemini 3.6 Flash scores 49% on DeepSWE versus 37% for 3.5 Flash, and 63.9% on MLE Bench versus 49.7%. Computer use performance on OSWorld-Verified improves from 78.4% to 83%, and knowledge work scores on GDPval-AA v2 rise from 1349 to 1421. The model carries a knowledge cutoff of March 2026 and ships with enhanced Frontier Safety safeguards covering chemical, biological, radiological, nuclear, and cyber offense domains, with training to minimize refusals for beneficial uses. It is a proprietary, closed-weights model available in preview through the Gemini API via Google AI Studio, Android Studio, Google Antigravity, Gemini Enterprise, and the Gemini app.

Kimi K3 is a sparse Mixture-of-Experts large language model developed by Moonshot AI, with 2.8 trillion total parameters and a 1-million-token context window. The model activates 16 out of 896 experts per token using the Stable LatentMoE framework, and is built on two architectural innovations: Kimi Delta Attention (KDA), a hybrid linear attention mechanism that enables up to 6.3x faster decoding in long-context settings, and Attention Residuals (AttnRes), which selectively retrieves representations across model depth and delivers roughly 25% higher training efficiency. Together with refined training and data recipes, these structural advances yield approximately 2.5x better overall scaling efficiency compared to its predecessor Kimi K2. The model applies quantization-aware training from the supervised fine-tuning stage onward, using MXFP4 weights with MXFP8 activations for hardware compatibility. Thinking mode is always enabled at launch, with reasoning effort configurable via the reasoning_effort field.Kimi K3 supports native visual understanding alongside text, accepting image inputs for tasks that combine software engineering and visual reasoning. It targets long-horizon coding, knowledge work, and agentic workflows, and ships in two variants: K3 Max for general chat and agent tasks, and K3 Swarm Max for large-scale parallel processing across many coordinated sub-agents. The model is compatible with the OpenAI SDK via an OpenAI-compatible API. Full model weights are scheduled for release by July 27, 2026 under a Modified MIT license, following the open-weight pattern established by the Kimi K2 model family. A technical report with full architecture, training, and evaluation details is expected to accompany the weights release.

GPT-5.6 Luna is the fastest and most cost-efficient model in OpenAI's GPT-5.6 family, which also includes Sol (the flagship tier) and Terra (the balanced mid-tier). Introduced under a new naming convention where the generation number (5.6) and a durable capability tier name (Luna, Terra, Sol) together define each model, Luna occupies the lightweight end of the family and is designed for high-volume, latency-sensitive workloads such as summarization, drafting, autocomplete, classification, and routine automation. The GPT-5.6 family as a whole advances capabilities in software engineering, computer use, professional knowledge work, scientific research, and cybersecurity, with all three tiers rated at the "High" capability level under OpenAI's Preparedness Framework for both cybersecurity and biological/chemical risk domains.GPT-5.6 Luna supports multimodal input and function calling, and shares the family's 1.5 million token context window. On Terminal-Bench 2.1, Luna scores 82.5%, and on the Artificial Analysis Coding Agent Index it outperforms comparable models at roughly one-quarter the estimated cost of higher-tier alternatives. Luna is priced at $1 per million input tokens and $6 per million output tokens, with cached input reads at $0.10 per million tokens under the GPT-5.6 prompt caching scheme, which introduces explicit cache breakpoints and a 30-minute minimum cache life. The model was previewed on June 26, 2026 to a limited group of trusted partners via the OpenAI API and Codex, with general availability rolling out on July 9, 2026 across ChatGPT, Codex, and the API.

GPT-5.6 Sol is the flagship model in OpenAI's GPT-5.6 family, which also includes Terra (a balanced everyday-work tier) and Luna (a fast, cost-efficient tier). Sol is designed for demanding reasoning, long-horizon agentic workflows, software engineering, computer use, scientific research, and cybersecurity tasks. It introduces two new capability modes: a "max" reasoning effort setting that allocates additional compute time for difficult problems, and an "ultra" mode that coordinates multiple subagents in parallel to accelerate complex, multi-step work. The model supports native multimodal input, allowing it to process screenshots, diagrams, charts, documents, and photographs alongside text. A reported context window of approximately 1.5 million tokens enables processing of large codebases, lengthy research documents, and extended agentic sessions.GPT-5.6 Sol was announced on June 26, 2026, initially in a limited preview for trusted partners, and reached general availability on July 9, 2026. On the Agents' Last Exam benchmark, which evaluates long-running professional workflows across 55 fields, Sol scores 53.6. On Terminal-Bench 2.1, which tests command-line agentic coding workflows, Sol Ultra achieves 91.9%. The model also demonstrates gains in life sciences evaluations, including long-horizon genomics and quantitative biology analyses. OpenAI paired the release with its most extensive safety evaluation to date, combining human red teaming with large-scale automated testing, and classified Sol as High capability in both cybersecurity and biological risk under its Preparedness Framework, though it does not cross the Critical threshold in either category.

GPT-5.6 Terra is the mid-tier reasoning model in OpenAI's GPT-5.6 family, which also includes the flagship Sol and the lightweight Luna. Introduced in a limited preview on June 26, 2026, and made broadly available on July 9, 2026, Terra accepts text and image input and produces text output, supporting vision, function calling, tool use, and agentic workflows. It is designed as a balanced option for everyday professional and production workloads — including coding assistance, document analysis, customer support, and multi-step agent tasks — where both output quality and cost efficiency matter. OpenAI positions Terra as delivering performance competitive with GPT-5.5 at approximately half the price, with a context window of around 1,050,000 tokens. On Terminal-Bench 2.1, Terra scores 84.3%, matching Claude Fable 5 on that benchmark. Under OpenAI's Preparedness Framework, Terra is rated High for cybersecurity and biological capabilities, meaning it demonstrates meaningful capability in those domains without reaching the Critical threshold.GPT-5.6 introduces a new naming convention in which the generation number (5.6) is paired with a durable capability tier name (Sol, Terra, or Luna), allowing each tier to advance on its own schedule. Terra carries the API identifier gpt-5.6-terra and supports the same reasoning effort controls available across the family, including adjustable reasoning depth. The model includes prompt caching with explicit cache breakpoints and a 30-minute minimum cache life, with cache writes billed at 1.25x the uncached input rate and cache reads receiving a 90% discount. GPT-5.6 Terra is a proprietary, closed-weights model served through the OpenAI API, Codex, and ChatGPT.

Muse Spark 1.1 is a natively multimodal reasoning model from Meta Superintelligence Labs, released on July 9, 2026, as a significant upgrade to the original Muse Spark. The model accepts text, image, video, PDF, and audio as input and produces text output. It operates with a 1-million-token context window (1,048,576 tokens per the Meta Model API documentation) and is designed specifically for agentic tasks that require planning, tool use, computer use, and multi-agent orchestration. The model runs in a "Thinking" mode, where adjustable reasoning effort is applied before generating a response. It can function both as a main agent gathering context, forming plans, and delegating to parallel subagents and as a subagent that adheres to assigned tasks and escalates when needed. It is trained to decide autonomously when to write automation scripts versus interact directly with a user interface.Muse Spark 1.1 supports a range of multimodal capabilities including visual perception, image and video captioning, visual-to-code generation, and document analysis. The model was evaluated under Meta's Advanced AI Scaling Framework across frontier risk categories including chemical and biological threats, cybersecurity, and loss-of-control scenarios. Parameter count, architecture details, and training data composition are not publicly disclosed. The model is proprietary and closed-weight, accessible to consumers through the Meta AI app and to developers via the Meta Model API, which launched in public preview alongside this release.

Claude Sonnet 5 is a mid-tier large language model from Anthropic, released on June 30, 2026, as the latest model in the Sonnet series and a direct successor to Claude Sonnet 4.6. It is a hybrid reasoning model designed primarily for agentic workflows, software coding, and professional tasks. The model features a 1 million token context window, a 128k maximum output token limit, and runs adaptive thinking by default, giving API users fine-grained control over reasoning effort across five levels (low, medium, high, max, and extra-high). It uses an updated tokenizer shared with Opus 4.7 and later models, which produces approximately 30% more tokens for equivalent text compared to earlier Claude models. On benchmarks, Sonnet 5 scores 63.2% on agentic coding and 81.2% on OSWorld, narrowing the gap with Opus 4.8 while remaining at Sonnet-tier pricing.The model supports text and image input with text output, and accepts tools including browsers and terminals for autonomous multi-step task execution. Anthropic's safety evaluations report that Sonnet 5 shows a lower rate of undesirable behaviors than Sonnet 4.6 and is generally safer in agentic contexts, with improved resistance to prompt injection and reduced sycophancy. Cybersecurity safeguards equivalent to those on Opus 4.7 and 4.8 are active, though Anthropic notes the model was not deliberately trained on cybersecurity tasks. The model is proprietary and API-only, with no open weights.

Claude Fable 5 is Anthropic's first generally available Mythos-class large language model, released on June 9, 2026. It is built for long-horizon, asynchronous, and agentic tasks that prior Claude generations could not sustain, including multi-day autonomous coding sessions, complex knowledge work, and document-heavy analysis. The model supports a 1 million token context window with up to 128,000 output tokens per request and uses adaptive thinking as its sole reasoning mode, where the effort level is adjustable but raw chain-of-thought is never returned. Vision capabilities allow the model to parse diagrams, charts, and tables embedded in files and PDFs, and to use visual feedback to evaluate its own coding outputs against design goals. On benchmarks such as SWE-Bench Pro, the model scores 80.3% compared to 69.2% for Claude Opus 4.8, and it leads on CursorBench 3.1 for autonomous coding workflows.Claude Fable 5 shares the same underlying model weights as Claude Mythos 5, but is deployed with safety classifiers that automatically reroute queries in high-risk domains — including cybersecurity, biology, and chemistry — to Claude Opus 4.8. These classifiers trigger in fewer than 5% of sessions on average. As a designated Covered Model, all traffic is subject to mandatory 30-day data retention to support safety monitoring. The model is available via the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Anthropic has not publicly disclosed parameter count, architecture details, or training data composition for this model.

Claude Opus 4.8 is Anthropic's most capable generally available large language model, released on May 28, 2026 as an incremental upgrade to Claude Opus 4.7. The model accepts text and image inputs and produces text outputs, with a 1 million token context window on the Claude API, Amazon Bedrock, and Google Cloud Vertex AI (200k tokens on Microsoft Foundry) and up to 128k max output tokens. It uses adaptive thinking and supports adjustable effort tiers — high by default, with extra and max tiers available for more demanding tasks. A fast mode operates at approximately 2.5x standard speed. The model is described by Anthropic as a hybrid reasoning model designed for advanced coding, agentic workflows, long-context reasoning, and professional knowledge work.Key behavioral improvements over Opus 4.7 include substantially reduced rates of unreported code flaws, improved honesty in self-assessment, and better tool-calling reliability. On Anthropic's Super-Agent benchmark, Opus 4.8 completes every case end-to-end, and it scores 84% on Online-Mind2Web for computer-use and browser-agent tasks. It achieves 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro. Alongside the model, Anthropic launched Dynamic Workflows in Claude Code (research preview), which enables Claude to orchestrate hundreds of parallel subagents for codebase-scale tasks such as large migrations. The Messages API was also updated to accept mid-task system messages without breaking prompt caching, improving support for long-running agentic pipelines.

Gemini 3.5 Flash is a multimodal language model developed by Google DeepMind and released at Google I/O 2026. It is built on the Gemini 3 Flash reasoning foundation and introduces configurable thinking levels (minimal, low, medium, and high) that allow developers to tune the depth of internal reasoning before a response is generated. The model accepts text, image, video, audio, and PDF inputs and produces text output, with a 1 million token context window and up to 65,000 output tokens per request. It is natively multimodal, processing visual inputs alongside text to support tasks such as image captioning, classification, optical character recognition, object detection, and visual grounding, where the model references specific regions within an image or video frame.Its vision capabilities extend to interpreting UI screenshots, diagrams, charts, and real-world scenes, as well as understanding video and live frame sequences for activity and scene recognition. The model supports combined tool use, including Google Search, URL context, code execution, and custom functions, within a single request, and it uses reasoning context from previous turns when thought signatures are present in the conversation history, enabling persistent multi-turn reasoning chains. Gemini 3.5 Flash carries a knowledge cutoff of January 2026 and is available via the Gemini API, Google AI Studio, Google Antigravity, and the Gemini Enterprise Agent Platform.

GPT-5.5 is a multimodal large language model released by OpenAI on April 23, 2026, engineered for autonomous, multi-step knowledge work and agentic workflows. It accepts text, images, and code as input, featuring enhanced spatial reasoning and visual grounding to support its computer use capabilities for operating software and navigating UI elements. Built to execute complex workflows end-to-end, the model interprets loosely defined tasks, selects appropriate tools, and performs self-verification with minimal user intervention. It is available in a standard version, a Thinking mode for extended reasoning budgets, and a Pro variant that uses parallel test-time compute for maximum precision on complex tasks.Co-optimized with NVIDIA for GB200 NVL72 infrastructure, GPT-5.5 delivers per-token latency comparable to its predecessor GPT-5.4 while maintaining a 1-million-token context window. Despite increased capability, the model achieves greater token efficiency in coding and data analysis workflows, often completing tasks with fewer total tokens than previous versions. OpenAI reports a 60% reduction in hallucination rate compared to GPT-5.4, improving reliability for accuracy-sensitive applications. API access is available via the Responses and Chat Completions endpoints at $5 per million input tokens and $30 per million output tokens, double the unit price of GPT-5.4.

Claude Opus 4.7 is a proprietary multimodal language model developed by Anthropic, released on April 16, 2026. It is designed for agentic coding, long-horizon task execution, and enterprise knowledge work. The model supports text and vision inputs and operates with a context window of up to 1,000,000 tokens. It introduces adaptive thinking, which dynamically allocates reasoning based on task complexity, along with configurable effort controls including a new xhigh setting that sits between the existing high and max levels. It achieves 87.6% on SWE-bench Verified and 78.0% on OSWorld-Verified, reflecting strong performance on autonomous software engineering and computer use tasks respectively.Compared to Claude Opus 4.6, version 4.7 shows improved instruction following and higher reliability in extended agentic tasks. Vision capabilities now support high-resolution inputs up to 2,576px on the long edge (~3.75 megapixels), more than three times the resolution of prior Claude models, enabling finer interpretation of dense diagrams, UI screenshots, and document layouts. These improvements, combined with self-verification on long-running tasks and a new task budget system for controlling agentic loops, make it well-suited for complex software engineering, technical analysis, and multimodal vision workflows.

Qwen3.6-35B-A3B is a sparse Mixture-of-Experts (MoE) multimodal language model developed by the Qwen team at Alibaba Group. It carries 35 billion total parameters but activates only approximately 3 billion per forward pass via a learned routing mechanism, giving it the representational capacity of a large dense model at a fraction of the inference compute. The model is natively multimodal, processing images, documents, and video alongside text as a core architectural capability rather than an add-on. It supports a native context window of 262,144 tokens, extensible up to 1,010,000 tokens via YaRN. A key design feature is the unified thinking/non-thinking mode framework: users can switch between deliberate chain-of-thought reasoning and fast direct responses within a single model, and a "thinking preservation" option retains reasoning context across multi-turn agentic workflows to reduce redundant computation.The model is specifically optimized for agentic coding tasks, including repository-level reasoning, frontend workflow generation, multi-step tool use, and MCP (Model Context Protocol) integration. On SWE-bench Verified it scores 73.4%, on Terminal-Bench 2.0 it scores 51.5%, and on MCPMark it scores 37.0%. For vision-language tasks it achieves 92.0 on RefCOCO, 89.9 on OmniDocBench 1.5, and 83.7 on VideoMMMU. The model also supports Multi-Token Prediction (MTP) for speculative decoding. All Qwen3.6 open-weight models are released under the Apache 2.0 license.

Gemma 4 26B A4B is the Mixture-of-Experts variant in Google's Gemma 4 family, with 25.2B total parameters but only 3.8B active per token. Built from the same Gemini 3 research as the 31B dense sibling and released as open weights under the Apache 2.0 license, it supports a 256K token context window with text and image input and configurable thinking mode. The "A4B" in the name refers to its approximately 4B active parameters. The MoE design makes it significantly faster at inference than the dense 31B, running nearly as fast as a 4B-parameter model while delivering roughly 97% of the dense model's quality.For vision tasks, the 26B A4B shares the same multimodal capabilities as the 31B image understanding with variable aspect ratios and resolutions, and structured bounding box output for UI element detection. The tradeoff versus the 31B dense model is a small quality reduction in exchange for much faster inference and lower hardware requirements, fitting in 18GB of VRAM at 4-bit quantization. It ranked #6 among open models on the Arena AI text leaderboard at launch.

Gemma 4 31B is the largest dense model in Google's Gemma 4 family, built from the same research as Gemini 3 and released as open weights under the Apache 2.0 license. It supports a 256K token context window with text and image input, configurable thinking mode for step-by-step reasoning, and multilingual support across 140+ languages. The unquantized model fits on a single 80GB GPU.For vision tasks, Gemma 4 31B supports image understanding with variable aspect ratios and resolutions, and can output structured bounding boxes for UI element detection, making it useful for document parsing and UI understanding. Compared to Gemma 3, it delivers stronger reasoning and multimodal performance. It is part of a four-size family alongside the 26B A4B MoE variant and two on-device models (E2B, E4B), with the 31B dense variant optimized for output quality and fine-tuning over inference speed.

Qwen3.6 Plus is a flagship model in Alibaba’s Qwen Plus series, designed for agentic workflows, coding, and multi-step reasoning. It supports a 1 million token context window and up to 65,536 output tokens, with built-in reasoning capabilities. The model is available as a hosted, proprietary API through Alibaba Cloud. Compared to Qwen3.5, it improves reliability in multi-step execution and frontend code generation, with stronger performance on agentic coding tasks. It also supports document and image understanding, though its vision capabilities are more limited than dedicated Qwen-VL models. Qwen3.6 Plus is part of a broader Qwen ecosystem that includes both closed-source APIs and open-weight models.

GPT-5.4 mini is a fast, cost-efficient model developed by OpenAI and released on March 17, 2026, optimized for high-throughput workloads and subagent orchestration. It supports text and image inputs within a 400,000-token context window, making it ideal for processing extensive visual datasets and large codebases in a single request. Designed for low-latency production environments, the model integrates with key API features including function calling, web search, and tool-based computer use, allowing it to assist in automated workflows that require navigating digital interfaces.Compared to the previous GPT-5 mini, this version runs more than twice as fast while approaching the performance levels of the flagship GPT-5.4 on reasoning and coding benchmarks. While the larger GPT-5.4 introduces native, state-of-the-art computer-use capabilities, GPT-5.4 mini provides a scalable alternative for interpreting screenshots and reasoning over dense UI layouts. For vision tasks on Playground, it excels at extracting structured information from visual documents and assisting in agentic tasks that involve real-time interpretation of software interfaces alongside text.

GPT-5.4 nano is a high-throughput model developed by OpenAI and released on March 17, 2026, as the efficiency-optimized entry in the GPT-5.4 family. Engineered for cost-sensitive production environments and latency-critical workloads, it features an expanded 400,000-token context window that enables the processing of large document batches or extensive logs in a single pass. The model is primarily optimized for text-heavy operations, serving as a premier engine for high-volume classification, data extraction, ranking, and the orchestration of lightweight sub-agents where speed and low per-token costs are the primary requirements.While it supports text and image inputs, GPT-5.4 nano is designed as a text-first worker rather than a specialized visual reasoning tool. In multi-model architectures, it is best utilized for structured text tasks and simple coding sub-tasks, leaving intensive vision reasoning and UI navigation to its sibling, GPT-5.4 mini. Compared to the previous GPT-5 nano, this version provides a significant leap in reliability for structured outputs and tool calling, making it a dependable and economical choice for developers building scalable, automated pipelines that require rapid execution at the edge of the GPT-5.4 ecosystem.

GPT-5.4 is a proprietary multimodal large language model developed by OpenAI and released on March 5, 2026. It is designed for professional workloads such as advanced software development, research, and agentic automation. The model combines the general reasoning capabilities of the GPT-5 series with software engineering improvements derived from GPT-5.3-Codex. In the API and Codex environments it supports context windows of up to 1 million tokens, enabling long-context reasoning and large-scale code or document workflows.Compared with GPT-5.2, GPT-5.4 reduces false individual claims by 33% and lowers overall response errors by 18%, improving factual reliability across complex tasks. It is also the first general-purpose OpenAI release with native computer-use capabilities, allowing agents to interact with desktops, browsers, and external applications to complete multi-step workflows. The model family includes three variants: GPT-5.4 (standard), GPT-5.4 Pro for higher-performance workloads, and GPT-5.4 Thinking, a reasoning-oriented version in ChatGPT that presents an upfront plan before generating its response. The API also introduces a Tool Search system that allows models to retrieve tool definitions dynamically, reducing token usage in tool-heavy integrations.

Gemini 3.1 Flash-Lite is a natively multimodal reasoning model from Google DeepMind in the Gemini 3 series, based on the Gemini 3 Pro architecture. It processes text, image, video, audio, and PDF inputs within a 1 million token context window and produces text output up to 64K tokens. The model targets high-volume, latency-sensitive workloads and supports visual question answering, image and document data extraction, content moderation, classification, translation, automated speech recognition, and agentic data pipelines. It exposes configurable thinking levels of minimal, low, medium, and high, which set the depth of internal reasoning applied per request and let developers balance response quality against cost and latency.On benchmarks reported at launch, Gemini 3.1 Flash-Lite scores 86.9% on GPQA Diamond and 76.8% on the MMMU Pro multimodal benchmark, and reaches an Elo score of 1432 on the Arena.ai leaderboard. According to Artificial Analysis benchmarks, it produces a 2.5 times faster time to first answer token and a 45% increase in output speed relative to Gemini 2.5 Flash. It also shows improved instruction following, higher audio input quality for automated speech recognition tasks, and support for structured JSON output used in data extraction pipelines.

Qwen3.5-9B is a 9-billion-parameter multimodal foundation model developed by Alibaba Cloud's Qwen team, released on March 2, 2026 as part of the Qwen3.5 model family. Designed for efficient multimodal reasoning and long-context language tasks, it notably outperforms the older Qwen3-30B, a model more than three times its size, on key benchmarks including GPQA Diamond, IFEval, and LongBench.The model supports vision-language inputs through an early-fusion multimodal architecture built on a dense hybrid foundation of Gated Delta Networks and Gated Attention. It can also operate in a text-only mode by skipping the vision encoder during inference. It provides a 262,144-token context window (extensible to ~1M tokens via YaRN) and is released under the Apache License 2.0. Within the current AI landscape, Qwen3.5-9B offers a strong balance of capability and efficiency, making it well-suited for multimodal assistants, document analysis, long-context reasoning, and developer-deployed agentic systems.

Qwen3.5-122B-A10B is a high-capacity multimodal Mixture-of-Experts (MoE) model developed by Alibaba’s Qwen team as part of the Qwen3.5 model family. The architecture contains 122 billion total parameters while activating roughly 10 billion per token through sparse expert routing, allowing the model to balance large-scale reasoning ability with relatively efficient inference compared to dense models of similar size.The model is designed to process both text and visual inputs within a unified multimodal framework, enabling tasks that require reasoning across images, documents, charts, and natural language. This makes it suitable for applications such as document understanding, diagram interpretation, and complex visual question answering.Qwen3.5-122B-A10B supports a native context window of approximately 256,000 tokens, which can be extended further through techniques such as YaRN scaling to support very long-context workloads. Released under the Apache 2.0 license, it builds on earlier Qwen multimodal systems and provides developers with an open-weight model capable of handling demanding multimodal reasoning and analysis tasks.

Qwen3.5-27B is a multimodal dense hybrid model developed by Alibaba Cloud’s Qwen team and released in February 2026 as a high-precision entry in the Qwen3.5 "Medium" series. Unlike its Mixture-of-Experts (MoE) siblings, the 27B model utilizes a dense architecture combining Gated Delta Networks with a feed-forward structure, activating its full parameter suite for every inference to maximize reliability. This design provides the highest instruction-following and coding accuracy in its class, with a notable IFEval score of 95.0. The model features a native 262K-token context window, extensible to 1M tokens via YaRN (RoPE scaling), and is released under the Apache-2.0 license.Optimized for agentic workflows, Qwen3.5-27B employs an early-fusion architecture that treats visual and textual data as a unified stream for deep cross-modal reasoning. This unified approach allows the model to excel in technical analysis and software engineering, matching GPT-5-mini with a 72.4% score on SWE-bench Verified. While the larger MoE variants in the family lead in raw knowledge benchmarks, the 27B model offers a stable and high-density alternative for structured data extraction and spatial perception, contributing to the Qwen3.5 family’s generational leap in OCR accuracy over the previous Qwen3-VL series.

The Qwen3.5-35B-A3B is a native vision-language model developed by Alibaba Cloud’s Qwen team, released on February 24, 2026, as a high-efficiency entry in the Qwen 3.5 family. It utilizes a sophisticated hybrid architecture that integrates Gated Delta Networks with a sparse Mixture-of-Experts (MoE) system. While the model houses 35 billion total parameters, its routing mechanism activates only 8 routed experts and 1 shared expert per token, totaling approximately 3 billion active parameters. This design achieves cross-generational parity with the previous flagship Qwen3-235B dense model, delivering comparable reasoning and multimodal intelligence with significantly reduced inference latency and compute requirements. Available under the Apache 2.0 license, it is released in both base and instruction-tuned variants for seamless integration with open-source stacks like vLLM and Hugging Face Transformers.Designed for the emerging era of agentic AI, the model utilizes a unified multimodal foundation built through early-fusion training. This approach allows it to outperform the prior Qwen3-VL series in spatial grounding, document analysis, and UI/GUI interaction. It features a native context window of 262,144 tokens, which is extensible up to 1,010,000 tokensvia RoPE scaling, and provides global support for 201 languages and dialects. This combination of a compact active parameter count and frontier-level visual comprehension makes it a versatile tool for developers requiring a balance of high-throughput speed and sophisticated visual reasoning for long-context workflows.

Gemini 3.1 Pro is a proprietary multimodal model from Google’s Gemini 3 series, released in early 2026 and designed for advanced reasoning across large multimodal datasets. It accepts text, images, audio, video, and documents, supporting up to a 1-million-token input context with up to 64k output tokens. Compared with Gemini 3 Pro, it improves long-context synthesis and multi-step reasoning, enabling more reliable analysis of large documents, datasets, and software codebases.The model also advances visual understanding and grounding, allowing it to interpret UI screenshots, diagrams, and real-world scenes while referencing specific regions within images or video. These capabilities make Gemini 3.1 Pro well suited for multimodal workflows involving document processing, interface analysis, robotics research, and complex visual reasoning.

Claude Sonnet 4.6 is Anthropic's mid-tier large language model, released February 17, 2026, designed to balance performance, cost, and versatility for professional and developer use. It supports text and vision-based tasks with advanced reasoning, agentic capabilities, and Adaptive Thinking — a mode where the model dynamically scales its internal reasoning depth. A beta context window of up to 1,000,000 tokens (200K standard) enables processing of entire codebases or document collections in a single request. Parameters are undisclosed.Optimized for coding, computer use, long-context reasoning, agent planning, and knowledge work, Sonnet 4.6 delivers a full generational upgrade over Sonnet 4.5 and approaches Opus 4.5-level performance across many benchmarks at a fraction of the cost. It is the default model on Claude.ai, Claude Cowork, and is available via API and major cloud platforms — making it well suited for production workloads requiring strong reasoning without flagship pricing.

Qwen3.5-397B-A17B is a 397B-parameter (17B active) open-weight multimodal model developed by Alibaba’s Qwen team, released on 2026-02-16 under Apache-2.0. It supports text and image inputs with text outputs, combining a sparse Mixture-of-Experts architecture with Gated Delta Networks for efficient scaling. The model provides native vision-language reasoning and a large ~262K token context window, extendable to ~1M tokens.As the first open-weight release in the Qwen3.5 family, it positions itself as a high-capacity, long-context alternative in the large vision-language space, balancing scale and efficiency via sparse activation. It is designed for advanced reasoning, coding, agent workflows, and multimodal understanding tasks.

Claude Opus 4.6 is the flagship large language model from Anthropic, released on 2026-02-05 for advanced reasoning, complex coding, and enterprise agent workflows. It supports text and image inputs via API, offers a 200K-token standard context window with a 1M-token beta option, and enables outputs up to 128K tokens, with adaptive reasoning and context compaction for sustained tasks.As of 2026-02-17, Anthropic also released Claude Sonnet 4.6, extending the 1M-token context window to a broader tier. Opus remains positioned for maximum depth and benchmark performance, while Sonnet 4.6 brings long-context capability to more cost- and latency-sensitive production use cases.

Gemini 3 Flash is a proprietary multimodal large language model developed by Google through Google DeepMind, designed to deliver fast, cost-efficient reasoning across real-time products and developer workflows. Released in December 2025, it is the Flash-tier variant of the Gemini 3 family, balancing low latency with reasoning quality approaching Pro models.The model supports text, images, audio, and video, with an exceptionally large context window of roughly one million input tokens and outputs up to ~65k tokens. It emphasizes rapid responses for coding, summarization, analysis, and agentic tasks, and exposes configurable “thinking levels” via API to trade speed for deeper reasoning. Today, Gemini 3 Flash positions itself as a high-throughput, production-ready model, serving as the default in the Gemini app and Google Search’s AI Mode, optimized for scalable, interactive AI applications.

GPT-5.2 is OpenAI’s latest flagship large language model, released in December 2025. It is a proprietary, multimodal system supporting text and vision inputs, along with tool use, and features a 400,000-token context window designed for working with long documents, extended conversations, and complex workflows.Relative to GPT-5.1, GPT-5.2 is positioned by OpenAI as offering improved long-context reasoning, more capable tool use, and stronger performance on professional tasks such as writing, coding, spreadsheet work, and image interpretation. The model is available in multiple variants (including Instant, Thinking, and Pro) that balance speed, cost, and depth of reasoning, making GPT-5.2 a general-purpose model aimed at reliability and workflow robustness rather than minimal latency or lowest cost.

Claude Opus 4.5 is Anthropic’s most advanced large language model in the Claude Opus family, designed for high-end reasoning, coding, and autonomous agent workflows. Released in late 2025, it targets developers and enterprises that need reliable long-context understanding and strong multi-step problem solving in production environments.The model supports text and code natively, with reported multimodal capabilities for documents and images, and offers an exceptionally large context window of up to roughly 200,000 tokens. Claude Opus 4.5 emphasizes long-horizon task execution, complex code generation and refactoring, and sustained reasoning over large inputs. In the current landscape, it positions itself as a premium, accuracy- and reasoning-focused alternative to faster or cheaper peers, trading cost for depth and contextual fidelity. Typical applications include advanced coding assistants, research analysis, agentic automation, and enterprise knowledge workflows deployed via Anthropic’s API or major cloud platforms.

Released on November 19th, 2025, Segment Anything 3 (SAM 3) is a zero-shot image segmentation model that “detects, segments, and tracks objects in images and videos based on concept prompts.” This model was developed by Meta as the third model in the Segment Anything series.

Unlike its previous SAM models (Segment Anything and Segment Anything 2), you can provide SAM 3 with the prompt “shipping container” and it will generate precise segmentation masks for all shipping containers in an image. SAM 3 generates segmentation masks that correspond to the location of the objects found with a text prompt.

GPT-5.1 is an OpenAI frontier-grade model in the GPT-5 series, offering stronger general-purpose reasoning, clearer long-form responses, and improved instruction following. It introduces two variants—Instant and Thinking—that dynamically adjust computational depth. Instant focuses on fast, conversational replies, while Thinking provides deeper, more thorough reasoning for complex tasks. In ChatGPT, GPT-5.1 also powers an Auto mode that switches between these variants automatically based on task difficulty.The model supports significantly expanded context windows: up to 16K/32K/128K tokens for Instant (depending on tier) and up to 196K tokens for Thinking on paid tiers. GPT-5.1 is also compatible with ChatGPT tools such as web search, file and image analysis, and multi-step workflows.GPT-5.1 includes enhanced tone and style controls, allowing responses to be tailored using presets like Friendly, Professional, or Efficient, along with fine-grained adjustments for warmth, brevity, and emoji usage. Designed for broad applications in research assistance, coding, analysis, and conversational agents, GPT-5.1 serves as OpenAI’s primary full-capability successor to GPT-5 across ChatGPT and API integrations.

Claude Haiku 4.5 is Anthropic’s lightweight model in the Claude 4.5 series, released in October 2025 under a proprietary license. Designed for speed and cost efficiency, it delivers near-frontier performance while maintaining Anthropic’s AI Safety Level 2 standard. Haiku 4.5 supports both text and multimodal (text and image) inputs, integrates tool use and extended reasoning, and features a 200,000 token context window, making it adept at handling long or complex workflows. Though the parameter count remains undisclosed, it achieves about 73.3% on SWE-bench Verified, reflecting strong coding and reasoning ability. Haiku 4.5 is ideal for developers and researchers seeking rapid, cost-effective model calls for analysis, coding, or multimodal understanding.

Qwen3 VL 8B Instruct is an open-weight multimodal vision-language model developed by Qwen / Alibaba Cloud as part of the Qwen3-VL series, designed for instruction-following tasks that combine text with visual inputs such as images and video. Released around October 2025 under the Apache-2.0 license, it targets developers who need capable multimodal reasoning without the scale or cost of very large models.The model contains roughly 8.8 billion dense parameters and supports text, image, and video understanding with strong spatial perception, visual reasoning, and emerging visual agent abilities such as GUI interaction. A standout feature is its native ~256K token context window, extendable to around 1M tokens, enabling long-document reading and extended video comprehension. In today’s landscape, it balances openness, long-context capacity, and solid multimodal performance against heavier proprietary models. Typical applications include multimodal assistants, document and video analysis, visual question answering, and research or product prototyping where transparency and deployability matter.

Qwen3 VL 30B A3B Instruct is an open-weight multimodal large language model developed by Alibaba as part of the Qwen family, built for instruction-following tasks that unify text generation with visual and video understanding. Released around October 2025 under the Apache-2.0 license, it targets efficient, high-fidelity vision-language reasoning across very long contexts.The model accepts text and image inputs and produces text outputs, with strong performance in OCR, spatial reasoning, long-video understanding, and agentic or GUI-centric visual tasks. It uses a Mixture-of-Experts (A3B) design with ~31.1B total parameters and ~3B active per token, paired with Qwen3-VL’s unified multimodal stack (including Interleaved-MRoPE and DeepStack fusion) to process text, images, and video in a single architecture. OCR support expands to 32 languages, enhancing document workflows. With a native ~262K token context window (extendable further), it stands out today for its balance of scale, efficiency, long-context support, and open accessibility in multimodal systems.

YOLO26 is a real-time object detection model developed by Ultralytics, released in October 2025. It introduces a native end-to-end, NMS-free architecture that eliminates the Non-Maximum Suppression post-processing step, reducing CPU latency by up to 43% for the Nano variant compared to NMS-dependent versions. The model incorporates the MuSGD optimizer and ProgLoss with STAL for improved training stability and small-object detection, and removes Distribution Focal Loss to ensure maximum compatibility with ONNX and TensorRT export targets.YOLO26 supports object detection, instance segmentation, pose estimation, and oriented bounding box detection within a unified framework, with model sizes available from Nano to Extra Large. Its NMS-free design makes it particularly well suited for deployment scenarios where post-processing overhead is a bottleneck, such as embedded systems and real-time edge inference pipelines.

Claude Sonnet 4.5, released by Anthropic in September 2025, is the company’s most advanced Sonnet-series model, built for high-performance reasoning, coding, and long-horizon agentic workflows. It is a multimodal system that accepts both text and images, with a 200,000-token context window designed for handling large documents and extended interactions. Anthropic highlights its improvements in reliability, reduced sycophancy, and alignment, making it suitable for sustained enterprise use.The model delivers strong results in coding and autonomous workflows, achieving 61.4% on the OSWorld benchmark and leading performance on SWE-bench Verified. It introduces infrastructure features such as a memory tool (beta), checkpointing for Claude Code, parallel tool use, and tighter integration with VS Code. Compared to Opus, which targets broader reasoning, Sonnet 4.5 is optimized for structured, long-duration tasks. Positioned against leading offerings from OpenAI and Google, it is aimed at enterprise automation, software engineering, and research-intensive applications.

Qwen3 VL 235B A22B Instruct is a flagship multimodal vision-language model developed by Qwen (Alibaba Cloud), designed for instruction-following tasks that combine advanced text generation with visual understanding. It serves as a high-end open-weight model for developers and researchers building multimodal AI systems that require strong reasoning, perception, and long-context capabilities.The model supports interleaved text and image inputs, very long context windows (up to roughly 256K tokens), and efficient inference through a mixture-of-experts architecture with about 22B active parameters out of 235B total. In today’s landscape, it competes with top-tier proprietary vision-language models while offering the advantages of open weights and flexible deployment. Typical applications include multimodal assistants, document and image analysis, visual reasoning, and large-context instruction-based workflows.

GPT-5, released by OpenAI in August 2025, is a multimodal large language model that advances beyond the GPT-4 family with a new “unified system” architecture. This design allows the model to dynamically choose between fast responses and extended reasoning depending on task complexity. It supports text, code, and images, alongside stronger tool use and agentic workflows, making it more adaptable for real-world problem solving. While its exact context window size is not disclosed, GPT-5 is optimized for long-horizon reasoning and multi-step tool chaining, indicating substantially expanded capacity over its predecessors.The release introduced specialized variants: GPT-5 Pro, offering extended reasoning for complex workflows, and GPT-5 Codex, optimized for advanced coding tasks such as large-scale refactoring and code review. GPT-5 shows benchmark gains in coding, biomedical reasoning, multimodal analysis, and scientific tasks. Developers also gain new controls, such as verbosity and personalization parameters, for greater steerability. With these improvements, GPT-5 positions itself as OpenAI’s most capable and versatile model, suited for enterprise automation, research, healthcare, and sophisticated coding environments.

GPT-5 Mini, released by OpenAI on August 7, 2025, is a mid-tier variant of the GPT-5 family that balances cost, speed, and capability. It is multimodal, supporting both text and image inputs, and offers a substantial input context window of ~400,000 tokens with output lengths up to ~128,000 tokens. While less powerful than the full GPT-5, it inherits its safety tuning, instruction-following improvements, and multimodal reasoning, making it a practical choice for developers who need large context handling without the expense of premium models.GPT-5 Mini is optimized for affordability while retaining strong reasoning performance. Benchmarks show it outperforming earlier models such as GPT-4o on many multimodal and medical VQA tasks, though it lags behind GPT-5 on the most complex problems. Ideal use cases include prototyping, scalable content generation, document analysis, and mid-range reasoning tasks where efficiency and context capacity matter more than top-tier accuracy.

GPT-5 Nano, released by OpenAI on August 7, 2025, is the smallest and most cost-efficient model in the GPT-5 family. Like its larger counterparts, it is multimodal—accepting text and images, supporting tool use, structured outputs, and reasoning—but it is optimized for speed, low latency, and affordability. It features input and output token limits of roughly 272K and 128K tokens respectively, enabling large-context processing even at its compact scale. Its knowledge cutoff is around May 2024, slightly earlier than the full GPT-5 model.GPT-5 Nano is well-suited for high-volume or cost-sensitive deployments such as mobile apps, embedded AI systems, or rapid-response APIs. While it offers less depth on complex reasoning and coding tasks compared to GPT-5 Mini or Pro, it retains core multimodal and agentic capabilities, making it an attractive option where efficiency and scale matter more than maximum performance.

Claude 4.1 Opus, released by Anthropic in August 2025, is the upgraded flagship of the Claude 4 family, building on Opus 4 with stronger reasoning and agentic capabilities. Like its predecessor, it is multimodal and optimized for text, code, and tool use, with support for large context windows suited to multi-file codebases, technical workflows, and long-horizon problem solving. On benchmarks, Opus 4.1 improves coding performance, reaching ~74.5% on SWE-Bench Verified compared to Opus 4’s ~72.5%. It demonstrates more precise debugging, refactoring, and orchestration of agentic tasks while maintaining similar safety and alignment safeguards. It is best suited for enterprise-scale software development, research automation, and advanced reasoning workflows where reliability and depth of analysis are critical.

Gemini 2.5 Flash-Lite, released for general availability on July 22, 2025, is the most cost-efficient model in the Gemini 2.5 family, designed for high-volume and latency-sensitive tasks. It is multimodal, supporting text, images, video, audio, and PDFs as inputs, with text as its primary output. The model handles up to 1 million input tokens and generates outputs up to 64K tokens, making it suitable for large-scale document or media processing at low cost. It is built on a Sparse Mixture-of-Experts architecture with native multimodal support, though exact parameter counts are undisclosed.Flash-Lite offers the lowest usage cost among Gemini 2.5 models. It introduces developer controls for “thinking mode,” allowing fine-tuning of reasoning depth vs. efficiency. It also integrates native tools such as code execution, search grounding, and URL context. While strong on translation, classification, coding, and general multimodal reasoning, it lacks support for image or audio generation in its stable release and is less capable than Gemini 2.5 Flash or Pro on complex reasoning-heavy workflows.

Gemini 2.5 Flash, released on June 17, 2025, is Google DeepMind’s production-ready, efficiency-focused model in the Gemini 2.5 family. It is multimodal, accepting text, images, video, and audio as inputs, with text as the primary output format. The model supports 1 million input tokens and up to 65K output tokens, enabling it to process very large contexts such as books, long video transcripts, or extensive datasets. Its training knowledge extends to January 2025.Designed as a price-performance leader, Gemini 2.5 Flash balances speed and reasoning power, making it suitable for everyday enterprise and developer use cases without the higher latency and cost of Pro models. It supports advanced workflows like function calling, code execution, search grounding, URL context ingestion, and structured outputs. While efficient and scalable, output length is still limited compared to its input capacity, and multimodal outputs (e.g. image or audio generation) remain restricted to specialized or preview variants.

Grok 4, released by xAI on July 9, 2025, is the fourth-generation model in the Grok family and the most advanced to date. It is multimodal, supporting text, vision, tool use, and real-time web search, with a reported 256,000-token context window for long-form reasoning and document analysis. Its training data extends through November 2024, making it the most up-to-date Grok model at launch.The lineup includes Grok 4 Generalist for broad tasks, Grok 4 Heavy for higher-capacity reasoning, and Grok 4 Code optimized for programming and debugging. A notable feature is its always-on “Think” mode, designed for deeper multi-step reasoning. While xAI has not disclosed parameter counts, Grok 4 is positioned to compete with frontier models like GPT-5 and Claude 4, balancing real-time knowledge via web integration with structured tool use. It is best suited for coding, complex reasoning, and multimodal AI assistants.

Florence-2, introduced by Microsoft Research at CVPR 2024, is an open-source vision-language foundation model designed to unify diverse computer vision tasks within a single sequence-to-sequence framework. Unlike traditional models that specialize in specific tasks, Florence-2 accepts both images and text prompts and outputs text for tasks such as captioning, object detection, segmentation, OCR, and region-based grounding. It comes in two sizes—Florence-2-base (~230M parameters) and Florence-2-large (~770M parameters)—and is trained on FLD-5B, a large dataset of ~126M images with ~5.4B annotations.The model demonstrates strong zero-shot and fine-tuned performance, often rivaling larger vision-language systems while remaining lightweight and efficient. Released under the MIT license, all weights are publicly available, making it accessible for fine-tuning and deployment in applications like VQA, content tagging, accessibility, and research. Florence-2’s compact design, versatility, and openness position it as a practical alternative to larger proprietary multimodal models.

Gemini 2.5 Pro, released on June 17, 2025, is Google DeepMind’s most capable model in the Gemini 2.5 family, optimized for deep reasoning, coding, and complex multimodal tasks. It accepts text, images, audio, video, and PDFs as input and outputs text. The model supports 1 million input tokens with an output capacity of up to 65K tokens, enabling large-scale comprehension of datasets, codebases, and technical documents. Its training knowledge extends to January 2025.Pro outperforms earlier Gemini 2.0 models across benchmarks, including agentic coding tasks where it achieved ~63.8% on SWE-Bench Verified. It supports structured outputs, function calling, code execution, search grounding, and URL context, making it well-suited for enterprise, STEM, and developer workflows. However, it does not currently support image or audio generation in its stable release, and its higher computational cost and latency make it less efficient than Flash or Flash-Lite. It is available via the Gemini API, Google AI Studio, and Vertex AI.

Llama 4 Maverick, introduced on April 5, 2025, is one of the first models in Meta’s Llama 4 family, designed as a natively multimodal model supporting text + image inputs with text outputs. It employs a Mixture-of-Experts (MoE) architecture with 128 experts, activating ~17B parameters per token out of a pool of ~400B total parameters. This design improves scalability, efficiency, and reasoning capacity. Maverick has a 1M-token context window, enabling it to handle large documents, extended conversations, and multimodal reasoning. Its knowledge cutoff is August 2024.The model is released under the Llama 4 Community License and comes in both base and instruction-tuned (“Instruct”) versions. Maverick is widely deployed via Hugging Face, Google Vertex AI, Amazon Bedrock, and Oracle Cloud, making it one of the most accessible large open-weight models. However, it outputs text only (no image/audio generation) and, while input capacity is huge, output limits are typically much smaller. The MoE design also raises hardware demands, as maintaining 128 experts requires significant compute resources, and Meta’s license introduces restrictions around commercial-scale use.

Llama 4 Scout, released on April 5, 2025, is one of Meta AI’s first Llama 4 multimodal models, alongside Maverick. It accepts text + image inputs and produces text outputs, with a knowledge cutoff of August 2024. Scout is notable for its extremely large context window of 10 million tokens, making it well-suited for analyzing very long documents, extended conversations, or large codebases.Architecturally, Scout uses a Mixture-of-Experts (MoE) system with 16 experts, activating ~17B parameters per inference from a pool of ~109B total parameters, balancing capacity with efficiency. It officially supports 12 languages (including English, Arabic, French, Hindi, and Spanish), while offering multimodal reasoning for images (captioning, Q&A, recognition). Meta highlights that Scout can run on a single Nvidia H100 GPU, making it more accessible than larger-scale Llama 4 models. However, its output token limit is far smaller than its 10M input window, image input support is still constrained, and license restrictions apply for large-scale commercial deployments.

RF-DETR is a real-time transformer-based object detection model developed by Roboflow, with code and weights first released in March 2025 under the Apache 2.0 license. It is the first real-time model to exceed 60 AP on the Microsoft COCO benchmark, built on a DINOv2 vision transformer backbone with weight-sharing neural architecture search used to identify accuracy-latency trade-offs. The full family spans six sizes from Nano (30.5M parameters, 384×384 input) to 2XL (126.9M parameters, 880×880 input), with the accompanying research paper accepted to ICLR 2026.RF-DETR is designed for strong domain adaptability, achieving state-of-the-art performance on RF100-VL, a benchmark measuring generalization to real-world object detection tasks across diverse domains. It is deployable through Roboflow Inference and supports fine-tuning on custom datasets, making it well suited for domain-specific applications with limited training data.

YOLOE (YOLO with Everything) is an open-vocabulary object detection and segmentation model developed by THU-MIG at Tsinghua University, released in March 2025 under the AGPL-3.0 license. It extends the YOLO architecture to support open-vocabulary detection through text and visual prompts, enabling the model to detect arbitrary object categories beyond a fixed training set without retraining. The design integrates prompt encoding directly into the YOLO framework while preserving real-time inference speed.YOLOE is evaluated on COCO and LVIS benchmarks and supports both closed-set and open-vocabulary detection modes. It is built on the Ultralytics codebase and maintains compatibility with standard YOLO training and export workflows. YOLOE is suited for applications requiring flexible, prompt-driven object detection where the target object vocabulary may change at inference time.

YOLOv12 is an attention-centric real-time object detection model developed by researchers at Tsinghua University, with the arXiv paper published in February 2025 under the AGPL-3.0 license. It introduces an Area Attention module that partitions feature maps into regions and applies self-attention within each region, reducing the quadratic complexity of full self-attention while capturing long-range dependencies. It also incorporates R-ELAN for improved feature aggregation and scaled residual connections for training stability.YOLOv12-L achieves 54.0% AP on COCO, while the YOLOv12-N variant achieves 40.5% mAP at 1.62ms latency on an NVIDIA T4 GPU. The model is built on the Ultralytics codebase, supporting detection, segmentation, and other standard YOLO tasks at competitive real-time speeds.

Qwen-VL-Max is a proprietary vision-language model developed by Alibaba’s QwenLM team. Released on February 1, 2025, it is the flagship offering in the Qwen-VL family and sits above the VL-Plus tier in capability.The model supports text and image inputs and provides a context window of up to 131,072 tokens (with a maximum input size of 129,024 tokens), according to Alibaba Cloud Model Studio. While the parameter count for VL-Max has not been publicly disclosed, the broader Qwen2.5-VL series includes open-weight models scaling up to 72B parameters.Qwen-VL-Max is optimized for advanced multimodal applications such as document parsing, visual reasoning, multilingual analysis, and structured data extraction. Unlike the open Qwen2.5-VL variants, VL-Max is not available as open weights.

Qwen2.5-VL-7B-Instruct is a 7-billion parameter vision-language model from Alibaba’s QwenLM team, released on January 26, 2025 under the Apache 2.0 license. It is the instruction-tuned variant of the 7B scale in the Qwen2.5-VL family, designed to process multimodal inputs such as text, images, charts, documents, and video. The model enables structured outputs—including JSON for structured content and bounding boxes for visual localization. Weights are publicly available on Hugging Face and GitHub, making it suitable for both research and applied multimodal use.

DEIM is a training framework for DETR-based object detection models released in December 2024 by researchers at Intellindust AI Lab, City University of Hong Kong, Great Bay University, and Hefei Normal University. It enhances existing real-time DETR architectures by improving the matcher used during training, enabling faster convergence and higher accuracy without modifying the inference architecture or adding computational overhead at deployment time. DEIM introduces two core techniques: Dense One-to-One (O2O) matching, which increases the number of positive matches per target, and Matchability-Aware Loss (MAL), which down-weights low-quality matches generated by the dense strategy. The paper was accepted at CVPR 2025.When integrated with RT-DETR and D-FINE, DEIM consistently improves performance while reducing training time by up to 50%. Applied to RT-DETRv2, it achieves 53.2% AP with a single day of training on an NVIDIA 4090 GPU. DEIM-enhanced models including DEIM-D-FINE-L and DEIM-D-FINE-X achieve 54.7% and 56.5% AP at 124 and 78 FPS respectively on an NVIDIA T4 GPU. DEIM is released under the Apache 2.0 license. A successor, DEIMv2, was released in September 2025, adding DINOv3-based backbones and introducing ultra-lightweight variants (Pico, Femto, and Atto) for edge deployment.

D-FINE is a real-time object detection model introduced in October 2024 by researchers at the University of Science and Technology of China. It builds on the DETR family of transformer-based detectors by reformulating bounding box regression as a Fine-grained Distribution Refinement task. Rather than predicting box coordinates directly, D-FINE iteratively refines probability distributions over coordinate offsets across decoder layers, which provides finer localization granularity without adding inference cost. The architecture also replaces the encoder's CSP blocks with GELAN modules and inserts a Target Gating Layer after the decoder's cross-attention to reduce representational entanglement across queries. A second contribution, Global Optimal Localization Self-Distillation, transfers localization knowledge from refined deeper-layer predictions back to earlier decoder layers through internal self-distillation.D-FINE is released in five model sizes (Nano, Small, Medium, Large, and X), with D-FINE-L achieving 54.0% AP on the Microsoft COCO benchmark at 124 FPS on an NVIDIA T4 GPU, and D-FINE-X reaching 55.8% AP at 78 FPS. Pretraining on the Objects365 dataset further improves accuracy to 57.1% AP for the L variant and 59.3% AP for the X variant. The paper was accepted at ICLR 2025 as a Spotlight. Code and pretrained weights are released under the Apache 2.0 license, making the model suitable for commercial use.

YOLO11 is an object detection and multi-task vision model developed by Ultralytics, released in September 2024 under the AGPL-3.0 license. It is the latest generation in the Ultralytics YOLO series and supports object detection, instance segmentation, image classification, pose estimation, and oriented bounding box detection within a single unified framework. YOLO11 introduces architectural refinements that improve accuracy while reducing parameter count compared to YOLOv8 at equivalent model sizes.YOLO11 is available in five model sizes from Nano to Extra Large and is deployable through the Ultralytics Python package, Roboflow Inference, and export formats including ONNX, TensorRT, and CoreML. It supports fine-tuning on custom datasets through the standard Ultralytics training API.

YOLOv10 is a real-time end-to-end object detection model developed by THU-MIG at Tsinghua University, released in May 2024 under the AGPL-3.0 license. It introduces consistent dual assignments during training — using both one-to-many and one-to-one label assignment strategies — to eliminate the need for non-maximum suppression at inference time while maintaining competitive accuracy. This end-to-end design reduces inference latency compared to NMS-dependent detectors at similar accuracy levels.YOLOv10-B achieves 52.7% AP on COCO with 46% lower latency than YOLOv9-C at comparable performance. The model is available in six sizes from Nano to Extra Large, built on the Ultralytics framework, and exportable to ONNX, TensorRT, and CoreML. YOLOv10 is suited for latency-sensitive deployment scenarios where post-processing overhead is a constraint.

YOLO-World v2 Small (YOLO-World-S-v2) is the smallest variant of Tencent AI Lab’s YOLO-World v2 family, released around February 2024 under GPL-v3. With ~13 million parameters, it adopts a prompt-then-detect paradigm using offline vocabularies and is pretrained on large-scale datasets such as Objects365 and GoldG. The model processes image inputs at 640×640 or 1280×1280 resolutions and supports zero-shot open-vocabulary object detection, enabling recognition of novel categories from text prompts without retraining.Evaluations show competitive results across benchmarks like LVIS and COCO, while maintaining real-time efficiency. On an NVIDIA V100, the small variant reaches ~74 FPS at standard resolutions. Together with larger YOLO-World v2 models, it provides a scalable framework for efficient, open-vocabulary detection across diverse deployment settings.

YOLOv9 is a real-time object detection model developed by Chien-Yao Wang and Hong-Yuan Mark Liao at Academia Sinica, released in February 2024 under the GPL-3.0 license. It introduces Programmable Gradient Information (PGI), a mechanism that preserves complete input information through auxiliary reversible branches during training to address information loss in deep network layers. It also introduces the Generalized Efficient Layer Aggregation Network (GELAN), which achieves better parameter utilization compared to prior CSP-based designs.YOLOv9-C achieves 53.0% AP on COCO with 42% fewer parameters and 21% less computation than YOLOv8-C at comparable accuracy. YOLOv9-E achieves 55.6% AP. The model is deployable through Roboflow Inference and supports fine-tuning via the standard training pipeline in the official repository.

Moondream 2 is a small open-source vision-language model from Moondream, the company founded by Vikhyat Korrapati. It was first released in early 2024 and updated through mid-2025. At approximately 1.9 billion parameters, it is designed to run efficiently on consumer hardware such as laptops and edge devices while supporting a practical range of multimodal tasks. Moondream 2 combines a vision encoder based on SigLIP with a compact language backbone, trained for image understanding tasks rather than as a general chat model.The model accepts an image paired with a natural language prompt and produces text responses, supporting visual question answering, image captioning, and image-conditioned dialogue. Later Moondream 2 releases added object localization through a point API that returns coordinates for queried objects, along with improvements to OCR, counting, and document understanding. Moondream 2 is distributed under the Apache 2.0 license and is available through Hugging Face and the maintainer's distribution. Because the model is updated frequently, production deployments should pin to a specific revision rather than tracking the latest release. A successor model, Moondream 3 (Preview), was released in September 2025 with a 9B mixture-of-experts architecture and 2B active parameters, offering substantially stronger visual reasoning than Moondream 2 while retaining the efficiency-focused design. A referring expression segmentation extension to Moondream 3 was released in March 2026.

YOLO-NAS is an object detection model developed by Deci AI, released in May 2023 as part of the super-gradients open-source training library. The architecture was generated using Deci's proprietary Neural Architecture Search technology, AutoNAC, which searches for network structures that balance accuracy and inference latency on target hardware. This produced three model sizes (small, medium, and large) featuring quantization-friendly blocks that reduce accuracy loss when converting weights to INT8 precision for deployment on edge devices and mobile hardware.YOLO-NAS achieves competitive accuracy-latency tradeoffs against YOLOv5, YOLOv6, YOLOv7, and YOLOv8 on the Microsoft COCO benchmark at release, and ships with pretraining on Objects365 in addition to COCO. Note that YOLO-NAS uses a custom license: the surrounding super-gradients framework code is Apache-2.0, but the YOLO-NAS model weights are released under a separate non-commercial license that restricts production and commercial use. Teams evaluating YOLO-NAS for commercial applications should review the LICENSE.YOLONAS.md terms directly. Deci AI was acquired by NVIDIA in April 2024, and the super-gradients repository is no longer actively maintained by the original team. Users can still download and use the released weights, but no further updates or new variants are expected.

RT-DETR (Real-Time Detection Transformer) is an object detection model developed by Baidu, released in April 2023 under the Apache 2.0 license. It is the first transformer-based real-time object detector, addressing the inference speed limitations of earlier DETR models through an efficient hybrid encoder that decouples intra-scale interaction and cross-scale fusion, enabling the model to process multi-scale features without the high computational overhead of standard transformer encoders.RT-DETR achieves 53.1% AP on COCO at 108 FPS on an NVIDIA T4 GPU for the RT-DETR-L variant, outperforming comparably sized YOLO detectors at similar speeds. It maintains end-to-end inference without non-maximum suppression, simplifying deployment pipelines. RT-DETR established the baseline for real-time transformer detection and has been extended by subsequent works including RF-DETR and RT-DETRv2.

Grounding DINO is an open-vocabulary object detection model developed by IDEA Research, released in March 2023 under the Apache 2.0 license. It extends the DINO transformer-based detector with grounded pre-training, enabling it to detect arbitrary objects described by free-form text queries rather than a fixed set of predefined categories. The model integrates a text encoder with a visual backbone through a feature fusion module that aligns language and visual representations at multiple scales.Grounding DINO achieves strong zero-shot detection performance on COCO, LVIS, and ODinW benchmarks, and supports referring expression comprehension tasks. It is widely used as a foundation for open-vocabulary detection pipelines and as the detection backbone in systems such as Grounded-SAM. The model is particularly suited for applications requiring flexible, text-driven object localization across diverse domains.

YOLOv8 is an object detection and multi-task vision model developed by Ultralytics, released in January 2023 under the AGPL-3.0 license. It succeeds YOLOv5 and introduces an anchor-free detection head, a new C2f module for improved gradient flow, and a decoupled head that separates classification and regression tasks. These changes improve both accuracy and training efficiency compared to earlier Ultralytics models.YOLOv8 supports object detection, instance segmentation, image classification, pose estimation, and oriented bounding box detection within a unified codebase. It is available in five sizes from Nano to Extra Large and exports to ONNX, TensorRT, CoreML, and other formats. YOLOv8 is one of the most widely adopted detection models in production and is directly supported by Roboflow Inference for custom model training and deployment.

RTMDet is a real-time object detection model developed by OpenMMLab, released in December 2022 under the GPL-3.0 license. It adopts a single-stage detection architecture with large-kernel depthwise convolution in both the backbone and neck, enabling it to capture long-range spatial dependencies without the computational cost of full self-attention. The model family spans from RTMDet-tiny to RTMDet-x, covering a wide range of speed-accuracy operating points.RTMDet-x achieves 52.6% AP on COCO at 114 FPS on an NVIDIA 3090 GPU. The architecture supports instance segmentation and rotated object detection variants. RTMDet is included in the OpenMMLab ecosystem and is well suited for applications requiring fast, accurate detection with flexible model sizing.

Co-DETR (Co-Deformable-DETR) is an object detection model developed by researchers at Sense-X and OpenMMLab, released in November 2022. It improves upon standard DETR-based detectors by introducing a collaborative hybrid assignment training scheme that enables the encoder to learn from multiple auxiliary heads simultaneously, alongside the primary one-to-one assignment used during inference. This auxiliary supervision significantly accelerates convergence and improves overall detection accuracy without adding inference cost.Co-DETR is evaluated on the COCO benchmark, where it achieves 59.5% AP when applied to DINO-Deformable-DETR with a Swin-L backbone. With a ViT-L backbone it reaches 66.0% AP on COCO test-dev, outperforming prior methods at comparable model scales. It is suitable for high-accuracy object detection tasks where training efficiency and peak performance on standard benchmarks are priorities.

YOLOv7 is a real-time object detection model developed by Chien-Yao Wang and Hong-Yuan Mark Liao at Academia Sinica, released in July 2022 under the GPL-3.0 license. It introduces Extended Efficient Layer Aggregation Networks (E-ELAN) for improved gradient flow in the backbone, and trainable bag-of-freebies techniques including coarse-to-fine lead guided label assignment and auxiliary heads that improve accuracy without adding inference cost.YOLOv7 achieves 56.8% AP on COCO at 30 FPS on a V100 GPU at the time of release, establishing a strong accuracy-speed tradeoff among real-time detectors. It supports detection, instance segmentation, and pose estimation variants. YOLOv7 is deployable through Roboflow Inference and the standard training pipeline in the official repository.

OWL-ViT (Open-World Localization with Vision Transformers) is an open-vocabulary object detection model released in May 2022 by Google Research. It adapts a pretrained CLIP-style image-text model by removing the final pooling layer and attaching lightweight classification and box prediction heads to each Transformer output token, producing a detector capable of localizing arbitrary objects described by free-form text at inference time. Rather than being restricted to a fixed taxonomy such as the 80 categories in Microsoft COCO, OWL-ViT can detect object classes specified by a user's text query, including categories the model was never explicitly trained on.OWL-ViT accepts an image and a list of text queries as input, and produces bounding boxes with class assignments drawn from the supplied queries. It also supports one-shot image-conditioned detection, where a cropped image region is used as the query instead of text, allowing the model to find visually similar instances within a target scene. The model is released in multiple Vision Transformer sizes (ViT-B/32, ViT-B/16, ViT-L/14) and CLIP-pretrained variants, distributed through the Google Research scenic repository and Hugging Face under the Apache 2.0 license. A successor model, OWLv2, was released in June 2023, introducing the OWL-ST self-training recipe that scales training to over one billion pseudo-annotated examples and substantially improves detection performance on rare and long-tail categories while preserving the open-vocabulary interface.

YOLOX is an anchor-free object detection model developed by Megvii (Face++), released in July 2021 under the Apache 2.0 license. It applies anchor-free detection to the YOLO framework, decoupling the classification and regression heads to allow each to optimize independently, and introduces the SimOTA label assignment strategy for improved training convergence. YOLOX achieves strong accuracy-speed tradeoffs and outperforms YOLOv5 on COCO at comparable model sizes.YOLOX-L achieves 50.0% AP on COCO at 68.9 FPS on an NVIDIA V100 GPU. The model is available in a range of sizes from YOLOX-Nano to YOLOX-X and supports deployment through ONNX, TensorRT, and other standard export formats. It is suitable for real-time object detection applications and has been widely adopted in industrial and research detection pipelines.

YOLOS (You Only Look at One Sequence) is a transformer-based object detection model widely distributed through Hugging Face Transformers, released in June 2021 under the MIT license. It applies a minimally adapted Vision Transformer to object detection by representing both the image and detection tokens as a flat sequence processed by standard multi-head self-attention, without convolutional components or feature pyramid networks. The architecture demonstrates that detection can be performed without region proposals or multi-scale feature fusion.YOLOS achieves moderate performance on COCO relative to purpose-built detectors, with its primary contribution being a demonstration of the transferability of ViT pre-training to detection tasks. It is most appropriate for research contexts exploring transformer-based detection architectures and for scenarios where architectural simplicity is preferred over peak accuracy.

YOLOv4-tiny is a lightweight variant of YOLOv4 developed by Academia Sinica, released in November 2020. It retains the core YOLOv4 design principles while significantly reducing the number of convolutional layers and feature map channels to produce a model suitable for inference on devices with limited compute, including embedded hardware and mobile CPUs. It uses a simplified CSP backbone with fewer layers and two detection scales rather than three.YOLOv4-tiny is optimized for scenarios where inference speed is prioritized over peak accuracy, achieving substantially higher FPS than full YOLOv4 at the cost of reduced AP on standard benchmarks. It is commonly used in robotics, embedded vision systems, and applications where real-time detection is required without GPU acceleration.

DETR (Detection Transformer) is an end-to-end object detection model developed by Facebook Research (Meta), released in May 2020. It is one of the first models to eliminate hand-crafted components such as anchor generation and non-maximum suppression by framing object detection as a direct set prediction problem, solved with a transformer encoder-decoder architecture built on top of a CNN backbone.DETR achieves 42.0% AP on the COCO benchmark with a ResNet-50 backbone, performing comparably to a well-tuned Faster R-CNN at the time of release. Its attention-based design allows it to reason about global context and long-range dependencies within an image. DETR is primarily used as a research baseline and architectural reference, with subsequent works such as Deformable DETR and DINO building on its foundations to address its slower training convergence and limited small-object detection capability.

YOLOv4 is an object detection model developed by Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao at Academia Sinica, released in April 2020 via the Darknet framework. It combines a CSPDarknet53 backbone, PANet neck, and YOLOv3 detection head with a large set of training improvements — Bag of Freebies and Bag of Specials — that improve accuracy with minimal inference cost increase.YOLOv4 achieves 43.5% AP on COCO at 65 FPS on a Tesla V100 GPU. The Darknet implementation is the original version, distinguishing it from subsequent PyTorch-based reimplementations. It remains a widely referenced detection architecture and a supported training target in Roboflow Inference.

YOLOv5 is an object detection model developed by Ultralytics, released in June 2020 under the AGPL-3.0 license. It is implemented in PyTorch and introduced a more accessible and well-documented YOLO implementation compared to earlier Darknet-based versions, with an integrated training and export pipeline supporting a wide range of deployment targets. YOLOv5 uses a CSP backbone, PANet neck, and a single-stage detection head with anchor-based regression.YOLOv5 is available in five sizes from Nano to Extra Large and supports export to ONNX, TensorRT, CoreML, and other formats. It is one of the most widely deployed object detection models in production environments and remains a common starting point for custom detection model training due to its documentation, community support, and compatibility with Roboflow Inference.

EfficientDet is an object detection model developed by Google Research, released in November 2019. It introduces a compound scaling method that uniformly scales the resolution, depth, and width of the detection network, building on the EfficientNet backbone and a bidirectional feature pyramid network (BiFPN) for multi-scale feature fusion. This design achieves strong accuracy-efficiency tradeoffs across a family of models ranging from EfficientDet-D0 to D7.EfficientDet-D7 achieves 55.1% AP on COCO while remaining significantly smaller in parameter count than comparable models at the time of release. The model family is well suited for deployment scenarios where compute budget varies, as smaller variants can run on edge hardware while larger variants are competitive with heavier architectures on server-side inference.

Detectron2 is a computer vision model library developed by Facebook AI Research (Meta), released in September 2019. It serves as a comprehensive platform for object detection, instance segmentation, panoptic segmentation, keypoint detection, and DensePose, implemented in PyTorch. It is the successor to the original Detectron framework, which was written in Caffe2, and offers a more modular and extensible codebase designed for both research and production use.Detectron2 includes implementations of Faster R-CNN, Mask R-CNN, RetinaNet, Cascade R-CNN, Panoptic FPN, and several other architectures. Its modular design allows components such as backbones, necks, and heads to be swapped independently, making it widely used as a baseline framework in academic research. It supports training on COCO-format datasets and integrates with standard distributed training setups.

MediaPipe is an open-source framework developed by Google for building real-time machine learning pipelines across mobile, web, desktop, and edge platforms. First released in 2019, the framework uses a graph-based architecture where pre-built components called Calculators process streaming data such as images, video, and audio through configurable computation graphs. This design allows developers to compose perception pipelines from reusable building blocks without writing custom glue code between models. The current MediaPipe Tasks API replaces the earlier Solutions API and provides a unified cross-platform interface for vision, text, and audio.Rather than providing a single model, MediaPipe ships a suite of ready-to-use Tasks that wrap trained models for specific problems. These include MediaPipe Pose Landmarker for 33-point body landmark detection, Hand Landmarker for 21-point hand tracking, Face Landmarker which extends the earlier 468-point Face Mesh with blendshape outputs for facial expression, Selfie Segmentation for person-background separation, and Holistic Landmarker for combined body, hand, and face tracking. The Tasks prioritize on-device inference with low latency and support GPU acceleration where available, making the framework a common choice for mobile augmented reality, fitness and wellness applications, gesture-based interfaces, and accessibility features such as sign language recognition.

MobileNet SSD v2 is a lightweight object detection model developed by Google Research, released in January 2018. It combines the MobileNetV2 backbone with the Single Shot MultiBox Detector (SSD) framework to produce a model optimized for inference on mobile and edge devices. MobileNetV2 introduces inverted residuals and linear bottlenecks to reduce computation while maintaining representational capacity compared to its predecessor.MobileNet SSD v2 is designed for real-time on-device detection, making it suitable for mobile apps, embedded systems, and IoT devices. It performs object detection across a fixed set of categories and can be fine-tuned on custom datasets. It trades peak accuracy for reduced inference cost and model size relative to larger two-stage detectors.

Mask R-CNN is an instance segmentation model developed by Facebook AI Research (Meta), released in October 2017. It extends Faster R-CNN by adding a parallel branch that predicts binary segmentation masks for each detected object, independent of the classification and bounding box regression branches. A key contribution is RoIAlign, which replaces RoIPool with bilinear interpolation to preserve spatial correspondence between features and input pixels, significantly improving mask quality.Mask R-CNN achieves strong performance on the COCO instance segmentation benchmark and supports keypoint detection as an additional output head. It remains a foundational architecture in instance segmentation and is available through Meta's Detectron2 framework. The model is most appropriate for tasks requiring pixel-level object delineation, such as medical imaging, autonomous driving, and industrial inspection.

Faster R-CNN is an object detection model introduced by Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun at Microsoft Research, published at NIPS in June 2015. It advances upon Fast R-CNN and R-CNN by introducing the Region Proposal Network (RPN), a fully convolutional network that shares features with the detection network and generates object proposals at negligible additional cost. This makes Faster R-CNN the first near-real-time deep learning object detector based on region proposals.Faster R-CNN achieves strong detection accuracy on PASCAL VOC and MS COCO at the time of release. It remains a widely referenced architecture in computer vision research and is available through Meta's Detectron2 framework as a maintained PyTorch implementation. It is most appropriate for offline or server-side inference tasks where accuracy is prioritized over latency, as its two-stage pipeline carries higher inference cost than single-stage detectors.