Google: Gemma 4 12B

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.

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.

Pixtral-12B is a vision-language model introduced by Mistral AI in September 2024 under the Apache 2.0 license, designed to process both text and images in a unified context. With ~12 billion parameters in its decoder and an additional ~400 million in a custom-trained vision encoder, it supports long-context reasoning up to 128k tokens and accepts multiple images per input. Its architecture is optimized for handling variable image sizes and aspect ratios, making it flexible for diverse multimodal tasks.As Mistral’s first VLM, Pixtral-12B delivers strong performance not only on image-text reasoning benchmarks but also in text-only applications, positioning it as a versatile alternative to models like GPT-4V and LLaVA. Its open availability via Hugging Face and major cloud providers such as Amazon Bedrock and SageMaker makes it accessible for research and production. Typical use cases include document analysis, visual QA, data extraction, and multimodal assistants requiring both textual and visual understanding.

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.

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.