Mistral Large 4 open weights are out: a 1 trillion parameter model, benchmarks the vendor graded itself

Mistral Large 4 open weights are now downloadable. Mistral released the weights of its roughly 1 trillion parameter, natively multimodal model on October 29, 2026, three weeks after the public preview, and says it ranks as "the best open-weight model for coding". Every benchmark in the announcement is Mistral's own, so the useful question is which numbers hold up and which are the vendor grading itself.

What shipped

Mistral's announcement describes Large 4, nicknamed "Le Chonk", as its flagship open-weight model, with about 1 trillion total parameters. The Hugging Face repository name, Mistral-Large-4-1T-A52B, suggests about 52 billion active parameters, but the announcement does not state that figure directly, and early preview coverage used a lower number. Treat the active count as unconfirmed until the model card says it plainly. The announcement does not give a context window.

The model has three reasoning modes, low, medium and high, with high as the default. It is available through Mistral Studio, the Mistral API and OpenRouter, as downloadable weights deployable with vLLM or SGLang, through NVIDIA-hosted endpoints and containers, and on Microsoft Foundry, which Mistral describes as the first cloud provider to offer it.

The earlier preview was covered in our piece on the 1 trillion parameter model with weights still weeks away. This is the follow-through on that promise, and it arrived inside the "end of October" window.

The benchmark table, sorted by who graded it

Mistral reports these results for Large 4:

• FrontierCode 2.0 (run by Cognition, private): 19.7%, the top score in Mistral's chart.
• Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3, using 47% fewer tokens per task on average.
• Terminal-Bench 4.0: 41%, up 13 points from the preview's 28%, with 30% fewer output tokens.
• Artificial Analysis Cyber Index: 51%.
• BrowseComp: 84.5%, up from 75% for an earlier checkpoint.
• Finch (FinWorkBench): 73.3%, second among open models only to Kimi K3.
• B3 Agent Security Benchmark: 97.1% attack resistance, up from 93.3% in the preview.

Two of those come from Artificial Analysis, a third-party evaluator, though Mistral selected which to show. The rest are benchmark results Mistral reports itself or through a partner on a private set. The improvement from preview to release (28% to 41% on Terminal-Bench, 75% to 84.5% on BrowseComp) is the most credible kind of claim, because it compares the model to itself on the same tests. The cross-model rankings are the weakest.

The broadest claim, "Mistral Large 4 ranks as the best open-weight model for coding", rests on a human-preference test. Mistral says raters on Surge Code preferred Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, placing it behind only Claude Opus 5.5 among the models tested. The models in that comparison were chosen by Mistral, and Kimi K3 beat it on Finch, so "best open-weight model" holds for coding preference, not for everything.

The license is not plain open source

Mistral calls the terms the Mistral Ecosystem Open-Weight license, a modified MIT license. Use, modification, fine-tuning and commercialization are allowed. Businesses that run a Model-as-a-Service offering need a separate commercial license from Mistral.

That carve-out matters. The Open Source Initiative's Open Source AI Definition 1.0 asks for more than weights: enough information about training data for a skilled person to build a substantially equivalent system, the complete training and inference code, and the parameters, each under OSI-approved terms. A license with a hosting restriction and no training-data disclosure does not meet that bar, which is why "open-weight" is the accurate label. Our explainer on open weights versus open source covers how to check any model against that definition.

Price and cyber safeguards

Mistral lists API pricing at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Those are vendor-listed prices as of October 29, 2026 and can change.

The cyber section is the most interesting part of the announcement, because a trillion-parameter model with downloadable weights can be fine-tuned by anyone. Mistral says the model refuses malicious cyber requests 96.6% of the time on average, which it calls the highest it measures among open models. It says the guardrails are designed to separate defensive from offensive work, and that it has seen the model refuse black-box attacks against live targets. It also concedes "a potential for uplift to conduct real attacks" and says it is modifying its guardrails accordingly. In its words, "Safety has improved alongside capability."

A caution applies to every open-weight safety number: refusal rates measure the model as shipped. Once the weights are public, a determined user can fine-tune the refusals away, and no vendor can recall the files. Mistral's answer is a separate cybersecurity variant, in preview only for vetted enterprise design partners such as Orange Cyberdefense, with testing limited to systems the customer owns or is explicitly authorized to test.

Our take

Large 4 is a real release with a real license, and the preview-to-release gains look genuine. It is not yet proof that an open model has caught the closed frontier: the headline ranking is a preference test, the efficiency claims are Mistral's, and we found no independent cyber-uplift study of the released weights. If you build on it, test it on your own tasks, read the license before you host it for others, and watch for an independent evaluator's run of the final weights. That is the result that would settle the "best open-weight coding model" claim.