Mistral’s Le Chonk weights are out, but its own chart doesn’t show a cyber lead

Mistral has released the final open weights of Le Chonk, the nickname for Mistral Large 4, a model with one trillion parameters. The French lab calls it the best open-weight model for agentic coding and cybersecurity, Wired reported.

Mistral’s own release post gives a more detailed picture than the headline claim, and in one respect a less flattering one.

A modified MIT license with one commercial catch

The weights sit on Hugging Face under the repository name Mistral-Large-4-1T-A52B, and Mistral says they can be run with vLLM or SGLang. The model is natively multimodal, mixes instruct and reasoning behavior in a mixture-of-experts design, and handles more than 160 languages. Mistral says it was trained from scratch on European infrastructure.

The license is the Mistral Ecosystem Open-Weight license, described as a modified MIT license. It allows free use, modification, fine-tuning and commercial use. The one condition: a business that offers the model as a hosted service to others needs a separate commercial license from Mistral.

For those who would rather not host it, the API costs $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens. The model is also on Mistral Studio, OpenRouter, Mistral Vibe Code, NVIDIA-hosted endpoints and Microsoft Foundry, which Mistral calls the first cloud provider to offer it. Three reasoning levels (low, medium and high, with high the default) let users trade depth for cost.

What Mistral’s benchmarks actually show

On coding, Mistral reports 19.7% on Cognition’s private FrontierCode 2.0, the top open-weight score in its chart. It also reports 41% on Terminal-Bench 4.0, up 13 points from the preview’s 28%, and 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task. In a blind human-preference test from Surge, it was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and placed behind only Claude Opus 5.5 among the models evaluated.

On cybersecurity, the number is 51% on the Artificial Analysis Cyber Index, which covers finding, reproducing and patching vulnerabilities without breaking functionality. Mistral’s own chart places MiMo-V2.6-Pro and DeepSeek-V4.1-Flash above it. So the claim holds more clearly for coding than for raw cyber capability. All of these figures are Mistral’s and have not been independently verified.

Mistral also reports a 96.6% average refusal rate on malicious cyber requests, across JailbreakBench, StrongREJECT and AgentHarm, which it says is the highest among open models it measures, and 97.1% on the B3 Agent Security Benchmark for attack resistance.

Open weights as a security argument, with a dual-use warning

Pierre Stock, Mistral’s VP of science, told reporters that the model “offers frontier capabilities for enterprises and governments to defend against an increasing threat level in cyber.” His case, as Wired relays it, is that downloaded weights cannot be switched off remotely, unlike proprietary models. He also described Chinese models as potential “threat vectors”, a view Wired sets against the more measured comments of an outside expert.

Rafe Pilling of the security firm Sophos told Wired the technology is plainly dual-use, since anything freely downloadable can be abused as well as used productively. He also said all advanced models can go rogue without careful supervision, whatever their origin.

Mistral concedes a version of that risk. After monitoring live traffic on the preview, it says the model refused black-box attacks on live targets but could still give potential uplift to real attackers, and that it is adjusting its guardrails. A specialized variant for penetration testing, vulnerability assessment and red-teaming is in preview for design partners such as Orange Cyberdefense, with vetted access limited to systems the user owns or is authorized to test. Mistral says it built the model with vendors and agencies including Cloudflare, CrowdStrike, Cisco, Fortinet, Zscaler and the EU agency ENISA.

Competition: Strong on agents, still behind US flagships overall

The wider picture is less clear-cut than the “leapfrogged China” framing. When the preview launched on October 6, The Register reported that Artificial Analysis placed it between DeepSeek V4.1 Flash and OpenAI’s entry-level GPT6 Luna on its overall intelligence leaderboard, a gap the outlet described as a significant disadvantage against the flagships of OpenAI and Anthropic. The same report said it beats Thinking Machines Lab’s Inkling, described as the most capable US open-weights model, by a wide margin, and that Mistral’s €3 billion Series D round funds this work.

Chinese labs keep moving too. Our weekly roundup tracked Kimi K3.1 Pro and MiniMax M3.1 Flash arriving only days ago, so any lead in open weights may be short. Mistral’s own FinWorkBench score of 73.3% is second to Kimi K3 among open models.

Mistral has been building out its product line around this model; our guide to Mistral’s assistant, now called Vibe, covers the consumer side.

What the sources leave open

Mistral lists no hardware requirements, and the Hugging Face repository name implies about 52 billion active parameters, a figure the release post does not state outright. Independent evaluations of the final weights are still to come. Developers who want to test the cybersecurity claims can now download the weights and do so themselves.