Top Document Question Answering Models

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.

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.

Qwen3.6-27B is a dense 27-billion-parameter multimodal language model developed by Alibaba's Qwen team and released on April 22, 2026. It combines a causal language model with an integrated vision encoder, supporting text, image, and video inputs natively. The architecture employs a hybrid attention design that interleaves Gated DeltaNet linear attention blocks with standard Gated Attention layers across 64 transformer layers with a hidden dimension of 5,120. Unlike Mixture-of-Experts variants in the Qwen3.6 family, all 27 billion parameters are active on every inference pass, simplifying deployment and quantization. The model supports a native context window of 262,144 tokens, extensible to approximately 1,010,000 tokens via YaRN scaling. It is released under the Apache 2.0 license with open weights available on Hugging Face and ModelScope.The model introduces two notable capabilities relative to prior Qwen releases: enhanced agentic coding support covering frontend workflows and repository-level reasoning, and a Thinking Preservation mechanism that retains chain-of-thought reasoning context across multi-turn conversation history to reduce redundant token generation in iterative agent sessions. It supports both a thinking mode for multi-step reasoning and a non-thinking mode for faster responses within a single model. On coding benchmarks, Qwen reports scores of 77.2 on SWE-bench Verified, 59.3 on Terminal-Bench 2.0, and 48.2 on SkillsBench. Vision capabilities include chart understanding (CharXiv RQ: 78.4), OCR (CC-OCR: 81.2), and video understanding (VideoMME with subtitles: 87.7).

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.

Qwen3.6-Flash is the production API variant of the Qwen3.6 model series, developed by the Qwen team at Alibaba Group. It is built on the Qwen3.6-35B-A3B architecture, which combines a hybrid linear attention mechanism with sparse Mixture-of-Experts (MoE) routing to achieve high-throughput inference with reduced latency. The model is natively multimodal, processing both text and images within a unified early-fusion architecture, and supports 201 languages and dialects. It operates in a hybrid thinking mode, capable of generating explicit chain-of-thought reasoning before producing a final response, with the option to disable thinking for direct output. A Thinking Preservation feature allows reasoning context to be retained across multi-turn conversations, which is particularly useful for iterative agentic workflows.The model is trained with reinforcement learning scaled across large-scale agent environments and covers a broad range of tasks including agentic coding, frontend development, visual understanding, document processing, and tool use. Compared to the open-weight Qwen3.6-35B-A3B, the Flash API variant extends the default context window to 1 million tokens and includes built-in production features such as native function calling and official tool integrations. The underlying architecture achieves near-100% multimodal training efficiency relative to text-only training, and the model demonstrates strong performance on agentic coding benchmarks including SWE-bench Verified.

GLM-OCR is a multimodal OCR model for complex document understanding, built on the GLM-V encoder-decoder architecture by Zhipu AI. The model combines a 0.4B-parameter CogViT visual encoder pre-trained on large-scale image-text data, a lightweight cross-modal connector with efficient token downsampling, and a 0.5B-parameter GLM language decoder, totaling 0.9B parameters. To address the inefficiency of standard autoregressive decoding in deterministic OCR tasks, GLM-OCR introduces a Multi-Token Prediction (MTP) mechanism that predicts multiple tokens per step, significantly improving decoding throughput while keeping memory overhead low through shared parameters. Training proceeds through four stages: visual encoder pretraining with MIM, CLIP, and distillation objectives; vision-language pretraining on document parsing, grounding, and VQA data; supervised fine-tuning on curated OCR datasets covering text, formula, table, and key information extraction; and full-task reinforcement learning to improve accuracy and structural consistency.At the system level, GLM-OCR adopts a two-stage pipeline in which PP-DocLayout-V3 first performs layout analysis, followed by parallel region-level recognition. This design enables robust handling of diverse document layouts including tables, formulas, and multi-column text. The model supports document parsing and targeted recognition tasks, producing structured outputs in Markdown, JSON, and LaTeX formats across more than 100 languages. On the OmniDocBench V1.5 benchmark, GLM-OCR scores 94.62, and achieves 94.0 on OCRBench and 96.5 on UniMERNet for formula recognition.

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.