On October 29, Mistral released the weights of Mistral Large 4, nicknamed “Le Chonk,” on Hugging Face, according to Mistral’s announcement. It has one trillion parameters, with 52 billion active at a time, and a one-million-token context window. Earlier this month only a preview existed, so its claims couldn’t be checked. Now anyone with enough GPUs can run it. This article looks at what can now be tested and what the published numbers show. I did not run the model myself. Benchmark figures below come from Mistral’s announcement and Hugging Face’s launch post unless stated otherwise, and none have been independently verified.
What was released
Mistral calls the model open-weight. It uses the Mistral Ecosystem Open-Weight License, a modified MIT license. You can use, change, fine-tune and sell products built on it. One condition applies: businesses that offer the model as a hosted service need a separate commercial license from Mistral. That is why some people say open-weight is not the same as open source. Running it needs data-center GPUs, and Mistral says its full technical report is still coming.
Per Mistral’s model documentation, its own API costs $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. Because the weights are public, other hosts set their own prices.
What Mistral claims
Mistral says three more weeks of reinforcement learning improved the model. All figures in this section come from Mistral’s announcement and are self-reported, unless stated otherwise.
Mistral reports 19.7% on FrontierCode 2.0, a private benchmark from Cognition. That figure comes from Mistral’s announcement and does not appear on Cognition’s own FrontierCode 2.0 leaderboard, whose October 27 snapshot, taken two days before the weights were released, listed Mistral Large 4 at 8.8%. Each leaderboard row pairs a model with an agent harness, so the two numbers may not be comparable. Artificial Analysis, which runs its own version of the benchmark, reports 28% for DeepSeek V4.1 Flash, the highest among the open-weight models it published. Treat Mistral’s claim of a lead among open models as unconfirmed.
Mistral reports 41% on Terminal-Bench 4.0, up 13 points from the preview’s self-reported 28%. This is Mistral’s own run, not a result from the Terminal-Bench maintainers. Mistral also reports 97.1% attack resistance on the B3 agent security benchmark, up from 93.3%. The source is Mistral’s announcement, not a B3 publication.
On Artificial Analysis’s Coding Agent Index, Mistral reports 48%, tied with GLM-5.3, while using 47% fewer tokens per task. This figure is taken from Mistral’s announcement, not from an Artificial Analysis report.
What its own charts show
The comparisons in this section are read from the benchmark charts in Mistral’s announcement, not from the benchmark owners’ own leaderboards.
The leads are narrow. In Mistral’s FrontierCode chart, DeepSeek V4.1 Flash scores 19.3% against Mistral’s 19.7%. A 0.4-point gap is small enough that run-to-run variation could erase it, and, as noted above, Artificial Analysis reports a much higher score for DeepSeek V4.1 Flash. The benchmark is private, so outsiders can’t rerun it. In Mistral’s Terminal-Bench 4.0 chart, GLM-5.3 scores 40% against Mistral’s 41%, a similar one-point margin. This is version 4.0, so don’t compare it with scores from older versions.
The cyber claim is mixed. Mistral’s model card says the model leads open models on security work, but Mistral’s own chart for the Artificial Analysis Cyber Index shows 51%, below MiMo-V2.6-Pro at 56% and DeepSeek-V4.1-Flash at 57%. Hugging Face’s launch post ranks the model fifth among open-weight models on Artificial Analysis’s overall Intelligence Index. That post also still cites 93.3% on B3, the preview score, rather than the new 97.1%. So “best open-weight model for coding” is a claim about coding, not everything.
Safety and why it matters
Mistral says its built-in guardrails refused 96.6% of malicious cyber requests, the highest among the open models it measured. That is Mistral’s own measurement. It also says that in live traffic it saw potential for uplift to real attacks and is changing its guardrails.
Public weights add another limit. Anyone can retrain a downloaded model, so refusals can’t be fully enforced. That is a general limit of open models, not a Mistral finding. For defenders, open weights mean control: they can audit the model and host it themselves.
What comes next
Watch for the technical report and independent runs on Terminal-Bench and the Artificial Analysis indexes. Mistral has invited the community to run its own tests. Until those results appear, treat the numbers as Mistral’s. Whether open models can keep leading in coding is still speculation.
