Open-Weight Models

First US-lab open-weight agentic coding model since Llama 4. 131K context, FP8 KV cache, mixed sliding-window/global attention 3:1 across 40 layers, native interleaved reasoning. SWE-bench Verified 68.2 / Pro 44.5 / Terminal-Bench 30.1; runs on a Mac with 36 GB RAM via Ollama.

Frontier-class two-variant drop, both with 1M-token context. Pro reportedly matches or exceeds Claude Opus 4.6 on coding and reasoning benchmarks at a fraction of the price; Flash trades capability for cheaper deployment. Both on Hugging Face under MIT.

First open-weights release from the rebuilt HunYuan stack. 256K context, 80 layers plus an MTP layer; reports SWE-bench Verified 74.4 and Terminal-Bench 2.0 54.4. Trained in under three months from cold-start to release.

Dense 27B variant of the Qwen3.6 family with 262K native context (extensible to ~1M). Targets agentic coding and frontend repo-level tasks; reportedly outperforms Qwen3.5-397B-A17B on coding benchmarks.

Sparse MoE optimized for inference cost — Ant claims ~86% fewer tokens consumed than peers on the Artificial Analysis index, with 340 tok/s prefill on 4xH20. Tuned for tool-use and SWE-bench Verified.

Trillion-parameter MoE with 384 experts, 256K context, and a 400M MoonViT vision encoder. Scores 58.6 on SWE-Bench Pro and 54.0 on HLE-Full with tools, and supports 300-sub-agent swarms over 4,000 steps.

Lightweight MoE Qwen3.6 variant with thinking-preservation across turns, designed for cheap agentic coding loops. Native 262K context; first Qwen3.6 open-weight drop.

Tiny on-device model trained on a 28T-token corpus using RL plus Liquid's non-transformer hybrid architecture. Targets phone-class deployment for assistants and tool-use.

Sparse 256-expert MoE with 200K context, tuned for multi-file code edits and agentic tool-use loops. Trended top-30 on HF within 48 hours of release.

Vision-language model with bounding-box prediction and multilingual prompt support, designed for sub-250ms edge inference on consumer hardware.

Open-weight visual web agent built on Molmo 2; operates a browser via screenshots. First fully-open agent to top closed systems on four web-navigation benchmarks; 30K human task trajectories shipped with the weights.

Tuning of GLM-5 for long-horizon coding agents; reported 58.4 on SWE-Bench Pro, ahead of GPT-5.4 and Claude Opus 4.6 on that benchmark. Trained on Huawei Ascend hardware.

First Gemma family shipped without a custom license. Multimodal incl. audio; 31B dense reports 89.2 on AIME 2026 and 80.0 on LiveCodeBench v6 — large jumps over Gemma 3 27B.

Mistral's first text-to-speech model. Nine languages, ~70ms latency, zero-shot voice cloning from 3 seconds of reference audio. Non-commercial weights only; commercial via Mistral API.

Unifies the previously separate Magistral (reasoning), Pixtral (vision), and Devstral (agentic coding) lines into one MoE with 256K context. Day-zero support across vLLM, llama.cpp, SGLang, and NIM.

Hybrid MoE with multi-token prediction and a four-experts-for-the-cost-of-one activation trick, claimed ~3x faster inference than peers. NVIDIA also published 10T tokens of training data.

Interleaves transformer attention with Gated DeltaNet (linear-RNN) layers; matches OLMo 3 MMLU using 49% fewer tokens, with ~75% better long-context throughput. Weights, data, and training logs all public.

Multimodal reasoning model built on Phi-4-Reasoning + SigLIP-2. Includes a learned policy for when to enter chain-of-thought versus answer directly, aimed at saving tokens on easy queries.

Self-hostable agent framework with three-layer memory and 118 skills, built atop existing open-weight LLMs. Listed because of architectural impact, not weights — reached 95K GitHub stars in 7 weeks.

Hybrid attention+convolution MoE designed to scale Liquid's edge architecture upward. Day-zero llama.cpp/vLLM/SGLang support; runs on a single consumer GPU with 4-bit quant.

Five-model release covering 70+ languages, including three regional fine-tunes (Africa/W. Asia, S. Asia, Asia-Pacific/Europe). Targets phone-class multilingual deployment.

Headline Qwen3.5 release: native multimodal, 201-language support, and the largest open-weight MoE Alibaba has shipped. Smaller dense variants (122B, 35B, 27B, 9B → 0.8B) followed Feb 24 and Mar 2.

First frontier-class model from a publicly traded Chinese AI company; trained entirely on Huawei Ascend chips. Replaced six weeks later by GLM-5.1.

Hybrid sparse-MoE coding model that activates only 3B per token; >70% on SWE-Bench Verified using SWE-Agent scaffolding, putting it in range of much larger dense coding models.

Distilled image generators from the FLUX.2 family targeting sub-second generation. The 4B variant is the first FLUX model under a fully permissive license; pairs with the open Apache 2.0 FLUX.2 VAE.