Mistral releases Large 4's open weights: Le Chonk lands on Hugging Face with sharper coding and cyber scores

On October 6, Mistral introduced Large 4 — nicknamed Le Chonk — as a public preview: a 1-trillion-parameter, natively multimodal mixture-of-experts model, with the weights promised for later in the month. Today, the Paris lab delivered the second half of that announcement. In a post titled “Mistral Large 4 (Le Chonk) is now open weight,” the company said it is releasing the weights of an improved version of the model, shaped by three more weeks of reinforcement learning and feedback from the preview.

That distinction matters. What is downloadable now is not simply the preview model with a download link attached. Mistral says the extra training moved the needle most where it is pitching the model hardest — agentic coding and defensive cybersecurity — while also making it cheaper to run. For a European lab positioning itself as the open alternative to American and Chinese frontier models, the release is the moment the preview's promise gets tested by anyone with enough hardware to run it.

What actually landed#

The checkpoint is published on Hugging Face under the name Mistral-Large-4-1T-A52B, and Mistral says it can be deployed on your own infrastructure with vLLM or SGLang. The architecture is a granular mixture-of-experts design: roughly 1 trillion total parameters, about 52 billion active per token (49 billion excluding embeddings and output layers), 128 routed experts with four active per token plus one shared expert, a 1-million-token context window, and a 1.6-billion-parameter vision encoder for text-and-image input.

Two serving details are aimed squarely at operators. The checkpoint ships with an embedded DSpark speculative-decoding draft head, which Mistral measured at up to 1.9× higher throughput than serving the model alone, and a quantized NVFP4 checkpoint is available for Blackwell GPUs. If you would rather not self-host, the same model is served as mistral-large-4 through Mistral Studio and the Mistral API, and through OpenRouter, Microsoft Foundry, and NVIDIA-hosted endpoints with an optimized NIM container. Mistral also points developers at its Vibe Code coding assistant to try the new model on their own repositories.

Coding: the headline claim#

Mistral's central claim is that Large 4 is now the best open-weight model for coding, and the numbers it cites are all vendor-reported, not independently reproduced — worth keeping in mind throughout. On Cognition's private FrontierCode 2.0 benchmark, which judges whether a repository maintainer would merge an agent's change, the model scores 19.7%, ahead of the listed open-weight models. On Terminal-Bench 4.0 it reaches 41%, a 13-point improvement on the preview and, Mistral says, the lead among open-weight models, while using 30% fewer output tokens. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average.

Human preference adds a second, softer signal: in Surge Code's blind pairwise evaluations cited by Mistral, the model was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated. Developers also get a new cost lever: three reasoning modes — low for simple tasks, medium for everyday work, and high as the default for the hardest problems — letting teams trade performance against inference cost instead of paying for maximum reasoning on every prompt.

Cyber defence — and its caveat#

The cybersecurity pitch is where Mistral is most ambitious, and most candid. The model scores 51% on the Artificial Analysis Cyber Index, which tests discovering, reproducing and patching vulnerabilities without breaking existing functionality. Its built-in guardrails are designed to separate legitimate defensive work from malicious offensive work: Mistral reports an average refusal rate of 96.6% on malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm, and 97.1% resistance on the B3 Agent Security Benchmark, up from 93.3% in the preview.

But the announcement also discloses something labs rarely volunteer. Monitoring live traffic since the preview, Mistral says it saw the model refuse black-box attacks against live targets — and still sees “potential for uplift to conduct real attacks,” so it is modifying its guardrails, with further updates promised in the coming months. A specialized variant for penetration testing, vulnerability assessment and red-teaming is going to vetted enterprise design partners such as Orange Cyberdefense, restricted to owned or explicitly authorized systems. The honest reading: this is a capable defensive tool whose offensive ceiling the vendor itself is still managing, not a solved problem.

Licence, price, and the fine print#

The weights are published under the Mistral Ecosystem Open-Weight license, a modified MIT license. In practice: free to use, modify, fine-tune and commercialize, with one condition — businesses operating a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially. The announcement does not spell out where ordinary SaaS wrapping ends and Model-as-a-Service begins, so teams building hosted products on top of the weights should read the license text itself before committing, rather than relying on the headline “open.”

On price, the new API rates are $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens — roughly half the preview's list rates of $1.36 and $4.18. The broader enterprise story also improved on Mistral's own measures: Finch (FinWorkBench) climbed to 73.3%, second among open models to Kimi K3, and BrowseComp reached 84.5%, up from a 75% supervised fine-tuning checkpoint. Against that, the model card's Artificial Analysis Intelligence Index v5 score of 26 trails the leading Chinese open-weight models on the broad aggregate — a reminder that coding and cyber leadership is not the same as general-purpose leadership.

What to watch#

Three things will decide whether this release lives up to its billing. First, independent reproduction: every benchmark above comes from Mistral or from evaluations it cites, and the community's own runs on the downloadable checkpoint — not the preview API — are the real test. Second, the license boundary: how Mistral enforces the Model-as-a-Service carve-out will shape whether startups treat Le Chonk as infrastructure or as a gated asset. Third, practical self-hosting: a trillion-parameter model, however token-efficient, still needs a serious cluster, so the NVFP4 checkpoint and the DSpark draft head deserve as much scrutiny as the benchmark table. The weights are out; the verdict is now the community's to deliver.

Sources: Mistral's October 29, 2026 announcement, “Mistral Large 4 (Le Chonk) is now open weight”, with coverage from Unite.AI and AI Weekly. Benchmark figures are as reported by Mistral and have not been independently verified. This article is a follow-up to our October 6 preview coverage, Mistral launches ‘Le Chonk’. Cover and inline images are AI-generated editorial illustrations.