Mistral Opens 1T 'Le Chonk' Large 4 Under Modified MIT License
TL;DR
- VentureBeat's independent Artificial Analysis benchmarks confirm ML4 leads all open-weight models outside China on coding tasks, with Mistral's own 63% lower cost-per-task figure holding up under third-party scrutiny.
- VentureBeat is the only source to flag ML4's factual accuracy at 26%, below several rivals, the sharpest gap Mistral's release materials do not surface.
- TechCrunch reports ML4 uses 47% fewer tokens than GLM-5.3, making token efficiency the primary cost argument against the leading Chinese open-weight alternative.
Mistral released the weights for Mistral Large 4, a trillion-parameter multimodal mixture-of-experts model it nicknames "Le Chonk," under its Ecosystem Open-Weight license, a modified MIT with one carve-out. The company's own announcement states the pitch directly: "We release the weights of Le Chonk, advancing the frontier of open models in agentic coding and cybersecurity."
The benchmark story is the sales argument. Mistral reports 19.7% on FrontierCode 2.0 and 41% on Terminal-Bench 4.0, a 13-point jump from the earlier preview release, along with 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens than competitors on the same task. On the security side, the model scores 51% on the Artificial Analysis Cyber Index, refuses 96.6% of malicious cyber requests, and resists 97.1% of attacks on the B3 Agent Security Benchmark. Enterprise numbers include 73.3% on FinWorkBench and 84.5% on BrowseComp.
The license is where the "open" claim gets qualified. Mistral says the model is free to use, modify, fine-tune and commercialize, with one condition: "businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially." The announcement does not define where SaaS wrapping ends and MaaS begins.
API access runs $0.70 per million input tokens, $0.07 per million for cached input, and $2.10 per million output. Weights sit on Hugging Face, listed as Mistral-Large-4-1T-A52B. The company also lists OpenRouter, NVIDIA endpoints, vLLM and SGLang at launch, and says the model was "developed and trained from scratch on European infrastructure" with native fluency in 160+ languages.
The release lands in a busy stretch for the lab: it is our eleventh Mistral story in 90 days, and comes a day after the open release of the Timestral family of time-series models.
What others are reporting
-
TechCrunch Read →
Leads with cybersecurity sovereignty and names the 47% token-efficiency advantage over GLM-5.3; VP Pierre Stock's 'can't be cut off' quote is the sharpest articulation of Mistral's sales pitch.
"AI should be a tool for empowerment, not dominance." — Arthur Mensch, Mistral CEO
-
VentureBeat Read →
Only source with independent Artificial Analysis benchmark verification; surfaces the 26% factual-accuracy gap and provides a full license analysis distinguishing self-deployment rights from cloud-resale restrictions.
"I don't want to get disconnected. I want to own my model." — Mistral Chief Revenue Officer
-
The Register Read →
Frames the 3,800 European Grace Blackwell GPUs and the €3B Series D as a geopolitical counterweight to US providers; critiques closed-model refusal policies on legitimate security research tasks.
"ML4 pairs top-tier cyber performance with open weights and self-deployment, giving organizations both the capability and the autonomy to run advanced security work."
-
The Next Web Read →
Quantifies RL gains (Terminal-Bench 4.0 from 28% to 41%), names automated penetration testing as the lead enterprise use case, and contextualizes $6B raised on 3,800 GPUs against rival training budgets.
"AI is going to be like electricity" and no company should depend on someone who can cut off access "from one day to another." — Pierre Stock, VP of Science
Shared on Bluesky by 1 AI expert
Originally reported by mistral.ai
Read the original article →Original headline: Mistral Opens 1T-Param MoE 'Le Chonk' Mistral Large 4 With Modified MIT License