We're releasing the weights of Mistral Large 4 (Le Chonk)
B · Useful, limited Mistral AI's own launch post for the open-weight release of Mistral Large 4; the licence terms and prices are primary facts, but every performance figure is a vendor-selected benchmark and most of the page is partner endorsements.
Summary based on the full text·Original: about 10 minutes to read
In short
Mistral AI released the weights of Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter multimodal model first shown in public preview earlier the same month. The company says three more weeks of reinforcement learning made it the leading open-weight model for coding and defensive security, with fewer tokens per task. The licence is a modified MIT licence that is free for commercial use except for companies offering the model as a hosted service. The post leans hard on European sovereignty, backed by quotes from security vendors, banks and cloud partners.
Key numbers
- 1-trillion-parametersize of Mistral Large 4 (Le Chonk), a natively multimodal model now released as open weights
- $0.70 / $2.10API price per million input and output tokens ($0.07 per million cached input)
- 48%score on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 at 47% fewer tokens per task (vendor claim)
- 41%Terminal-Bench 4.0 score, up 13 points from the public preview
- 96.6%average refusal rate on malicious cyber requests across three jailbreak benchmarks (vendor measurement)
As reported by the source.
English·Web page·opens mistral.ai
What it reports
- Coding (vendor figures): 19.7% on Cognition's private FrontierCode 2.0, described as the lead; 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 at 47% fewer tokens per task; 41% on Terminal-Bench 4.0, up 13 points from preview with 30% fewer output tokens. In Surge Code's blind pairwise tests it was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, behind only Claude Opus 5.5.
- Security: 51% on the Artificial Analysis Cyber Index; 96.6% average refusal on malicious cyber prompts from JailbreakBench, StrongREJECT and AgentHarm; 97.1% attack resistance on the B3 Agent Security Benchmark, up from 93.3% in preview.
- Mistral says live-traffic monitoring showed refusals of black-box attacks on live targets but "a potential for uplift to conduct real attacks", and that it is changing its guardrails. A specialised offensive-security variant goes only to vetted partners under contract.
- Other scores: Finch (finance spreadsheet work) 73.3%, second among open models to Kimi K3; BrowseComp 84.5%, from 75% at the starting checkpoint.
- Three reasoning modes trade cost for accuracy. Weights on Hugging Face for vLLM or SGLang; also via Mistral's API, OpenRouter, NVIDIA endpoints and Microsoft Foundry. The model card cites fluency in 160+ languages. A technical report is promised.
- Licence: Mistral Ecosystem Open-Weight licence, free to use, modify, fine-tune and commercialise; Model-as-a-Service providers need a separate commercial licence.
How they got it
Vendor-run or vendor-selected benchmarks, partly on third-party indices, with charts. No technical report yet; architecture details such as active parameters and context length are not stated in the post.
Limits
- All performance claims come from the vendor, with no independent replication cited.
- Comparisons are mostly to other open-weight models; frontier closed models appear only in the human-preference result.
- The post does not give context length, active parameters or hardware needs.
- Most of the text is partner testimonials.
Why it matters for companies
A trillion-parameter model with open weights and a permissive licence gives European and regulated firms a self-hosting option near the front of open models, with published API prices. The licence carve-out matters for anyone building a hosted AI product on top of it. The security claims cut both ways: Mistral itself reports possible uplift for real attacks, which is a reason to apply your own controls rather than rely on the model's refusals. Independent benchmarks and the promised technical report are needed before relying on the coding claims.
Summary based on the full text. Read the original for the full argument and data.Read the original on mistral.ai