Mistral Large 4 Goes Open Weight: Europe’s Flagship Joins the Open Race Against DeepSeek and GLM

Open weights on October 29, after the October preview

Mistral AI has released the open weights of Mistral Large 4, also called Le Chonk, its flagship trillion-parameter multimodal model, on October 29, 2026. The downloadable checkpoint lands on Hugging Face under the Mistral Ecosystem Open-Weight license, after a public API preview that began earlier in October. my ai daily covered that preview in “Mistral Large 4 enters public preview ahead of a weights release.”

Independent evaluator Artificial Analysis puts the high-reasoning configuration at 26 on the Intelligence Index v5—narrowly the strongest open-weights model developed outside China, while still trailing Chinese open leaders DeepSeek V4.1 Flash and GLM-5.3 by six points.

At a glance

Mistral’s docs put total parameters at 1.05 trillion and active parameters at about 52 billion. The Hugging Face model card states 49 billion active per token, and 52 billion including embeddings and output. Context is 1 million tokens. Modalities are text and image in, text out, with a Pixtral-style vision encoder. Those figures come from the official docs and the model card.

On Artificial Analysis Intelligence Index v5, the High configuration scores 26. The Preview scored 25. AA’s open-weights standing for this release is first among models developed outside China, and six points behind DeepSeek V4.1 Flash and GLM-5.3, both at 32. Versus the Preview, AA reports 39% fewer output tokens per Intelligence Index task, and an AA-Omniscience v1.1 hallucination rate moving from 41% to 28%.

The license is the Mistral Ecosystem Open-Weight license, which Artificial Analysis abbreviates MEOW. Mistral says the weights are free to use, fine-tune and commercialize, but a Model-as-a-Service offering needs a separate commercial license. API pricing on the Mistral OSS blog is $0.70 per million input tokens, $2.10 per million output tokens, and $0.07 for cached input.

What shipped

Mistral first unveiled Large 4 in public preview on October 6, 2026, promising weights by the end of the month. The OSS post frames the downloadable model as an improved checkpoint after additional reinforcement learning—especially for agentic coding and defensive cybersecurity—plus clearer reasoning-effort controls: low, medium and high, with high as the default.

Weights are published as mistralai/Mistral-Large-4-1T-A52B on Hugging Face in FP8, with an NVFP4 quantized sibling for Blackwell GPUs. Day-0 serving paths called out by Mistral and Hugging Face include vLLM, SGLang, Transformers, and llama.cpp / GGUF community builds. Hosted options include the Mistral API (mistral-large-4), OpenRouter, NVIDIA NIM and Microsoft Foundry.

Architecturally, Large 4 is a granular mixture of experts: roughly 128 routed experts with 4 active per token plus one shared expert, paired with sparse attention and a roughly 1.6 billion-parameter vision encoder, per the docs. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters—central to the company’s sovereign AI pitch.

Open-weights landscape: Mistral vs DeepSeek and GLM

The useful comparison is not whether the model beats every open model, but where Europe’s open flagship sits relative to China’s open leaders. Per Artificial Analysis, using the High or Max settings as published by AA: DeepSeek V4.1 Flash 1017 (Max), framed as China, scores 32 and co-leads open weights. GLM-5.3 (Max), from Z AI in China, also scores 32 and co-leads. Mistral Large 4 (High), from France, scores 26 and is the best open weights developed outside China in that read. Muse Spark Open (Xhigh), from Meta in the US, scores 25 and is the next non-China open competitor AA lists.

AA attributes most of the six-point gap to weaker coding and agentic knowledge-work scores—for example 15% on FrontierCode 2.0 versus 29% for DeepSeek Flash and 22% for GLM-5.3, and lower Elo on AA-Briefcase. On the Coding Agent Index, though, Mistral reports, and AA corroborates a near-tie, about 48%, effectively matching GLM-5.3 while using far fewer tokens. Mistral cites 47% fewer tokens per task on that index. AA separately notes roughly half the tokens of GLM-5.3’s Opencode run.

Read carefully: Mistral’s own blog emphasizes coding and cyber leadership among open weights. AA’s aggregate Intelligence Index still puts Chinese open models ahead. Both can be true at once—different scoreboards, different strengths.

Efficiency over the Preview

AA’s launch note stresses that the headline Intelligence Index gain is only +1, from 25 to 26. The bigger shifts are cost and reliability. Token efficiency is about 82,000 output tokens per Intelligence Index task versus about 135,000 for the Preview, which is 39% fewer tokens. The AA-Omniscience v1.1 hallucination rate is 28%, down from the Preview’s 41%. Accuracy on that eval is still relatively low at 26%, per AA.

The API list price was cut roughly in half to $0.70 input and $2.10 output. AA estimates $2.06 per Intelligence Index task, down about 63% from the Preview’s $5.54. Mistral’s blog similarly highlights Terminal-Bench 4.0 rising to 41%, up 13 points versus the Preview, with 30% fewer output tokens, and Coding Agent Index performance tied with GLM-5.3 at lower token spend.

License: open weights with a MaaS caveat

The license is the Mistral Ecosystem Open-Weight license—AA abbreviates it MEOW—described by Mistral as a modified MIT license. You may use, modify, fine-tune and commercialize the weights, with one hard edge case: operating a Model-as-a-Service offering requires a separate commercial license from Mistral. Self-hosting for internal use, or shipping a product that embeds the model without offering the model itself as a service, sits on the free side of that line as stated by Mistral. Teams building public LLM APIs should read the license text before shipping.

Bottom line

Mistral Large 4’s open-weight drop is Europe’s clearest answer yet to the Chinese open-weights surge: a 1.05-trillion-parameter model with about 52 billion active parameters, a 1-million-token context and a multimodal mixture of experts you can download, with strong official claims on coding agents and defensive cyber. AA confirms it as the top non-China open model on Intelligence Index v5, while it remains six points behind DeepSeek V4.1 Flash and GLM-5.3 on that aggregate. For builders, the practical story may be less “new #1 overall” and more sovereign, self-hostable frontier-class weights with better token economics than the Preview—as long as MaaS operators budget for Mistral’s separate commercial license.