Mistral 3 release — vision-capable 3B/8B/14B dense + Large 3 (675B total, 41B active MoE) under Apache 2.0

TL;DR

Mistral AI’s Dec 2, 2025 announcement of the Mistral 3 family: three dense, vision-capable “Ministral 3” models at 3B, 8B, and 14B parameters (each in base / instruct / reasoning variants), plus Mistral Large 3, a sparse mixture-of-experts model with 675B total and 41B active parameters. All models are vision-capable and released under Apache 2.0. Notable for Luma: the 3B is small enough to run in-browser via WebGPU as a multimodal model, and Mistral Large 3 is Mistral’s first MoE since Mixtral, trained from scratch on 3,000 H200 GPUs.

Key claims

Four models released together: dense 3B / 8B / 14B (Ministral 3 series) plus Mistral Large 3, a sparse MoE with 675B total / 41B active parameters [tweet body, corroborated by the mistral.ai/news/mistral-3 announcement].
All models are vision-capable (multimodal text + image) and released under Apache 2.0 [tweet body].
Each Ministral 3 size ships in three variants — base, instruct, and reasoning — covering all combinations of {3B,8B,14B} × {base, instruct, reasoning} with native multimodal and multilingual capabilities [announcement blog].
Mistral Large 3 is Mistral’s first mixture-of-experts model since the original Mixtral series, trained from scratch on 3,000 NVIDIA H200 GPUs [announcement blog].
Deployment variants distributed for Large 3 include BF16 base, instruct, FP8 (for H200/B200 single-node serving), NVFP4 (for 8× A100/H100 nodes), plus an EAGLE speculative-decoding speculator in FP8 [HF model card listing].

Method

This is a model release announcement, not a paper. The release packages a dense-and-sparse pair: a small dense series targeting edge / local deployment (Ministral 3, 3B–14B) and a single large sparse MoE at the frontier scale (Large 3, 675B/41B). Per the announcement, the MoE design is described as “granular” — an evolution of the Mixtral architecture — and all models share native vision and multilingual capabilities. The Ministral 3 line is fanned out across three training objectives (base / instruct / reasoning) so the same backbone parameters appear under three post-trained heads per size.

Results

Headline numbers from the announcement and corroborating coverage (no independent verification at filing time):

Ministral 3 reasoning variants at 14B reportedly score ~85% on AIME ‘25.
Mistral Large 3 reported #2 on LMArena’s non-reasoning open-source leaderboard at release, with top results on GPQA Diamond among open models, and claimed to outperform DeepSeek-V3 and Llama 3.1 405B on those benchmarks.
256K context window across the family.
Day-0 availability on Mistral AI Studio, Amazon Bedrock, Azure Foundry, Hugging Face, IBM WatsonX, Modal, OpenRouter, Fireworks, Unsloth AI, and Together AI.

These numbers come from the Mistral announcement and aggregator coverage; the tweet itself doesn’t include a benchmark table.

Why it’s interesting

This release is the strongest open-weights MoE datapoint from a Western lab since Mixtral, and it lands directly in the territory the wiki has been tracking on Open foundation-model releases — vision-capable, multi-variant, multi-backend day-0 packaging, all under Apache 2.0. The 675B/41B sparsity ratio (~6.1%) is in the same neighborhood as DeepSeek-V4 collection release (Flash + Pro, up to 1.6T) and the trinity of Chinese-lab MoE flagships (Trinity Large: An Open 400B Sparse MoE Model, Kimi K2: Open Agentic Intelligence); the news here is that Mistral has now matched that pattern, not that the pattern is new. The 3B vision-capable model running entirely in WebGPU is a separate datapoint — extending the “small enough to ship to the client” frontier beyond text-only models like LFM2 Technical Report.