Mistral ships open weights for Le Chonk, its 1T-parameter Mistral Large 4

Mistral released downloadable weights for Mistral Large 4, a 1T-parameter MoE model it says leads open-weight coding and defensive cyber benchmarks.

What happened

Mistral has released the weights of Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter natively multimodal model it put into public preview earlier this month. TechCrunch reports the French company is making the weights available on Hugging Face and via its API, letting businesses and governments download and deploy the system themselves.

It reports 19.7% on Cognition's private FrontierCode 2.0 benchmark, which assesses end-to-end code quality and whether a repository maintainer would merge the agent's change, and 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task. On Terminal-Bench 4.0 the score improved 13 percentage points to 41% compared with the public preview while cutting output tokens by 30%. The model now supports three reasoning modes. Mistral also cites 51% on the Artificial Analysis Cyber Index, a 96.6% average refusal rate on malicious cyber requests, 73.3% on Finch, 84.5% on BrowseComp, and 97.1% attack resistance on the B3 Agent Security Benchmark, up from 93.3% in preview.

The Hugging Face blog describes a Mixture-of-Experts model with about 1T total and 49B active parameters, a 1M token context window, text and image input, sparse attention with an attention indexer that selects earlier token positions, and a Pixtral vision encoder. It is released in two official variants, with GGUFs created by Unsloth and the llama.cpp team. Deployment paths include Mistral Studio, the Mistral API, OpenRouter, self-hosting with vLLM or SGLang, Microsoft Foundry, and NVIDIA-hosted endpoints. Mistral says a full technical report is coming soon.

Why it matters

The practical change is availability: teams can download the weights now rather than consume the model only through a preview endpoint. Mistral says organizations can inspect, customize, and deploy the model on their own infrastructure and use it without depending on Mistral to serve it through an API. Pierre Stock, Vice President of Science at Mistral, told TechCrunch that access "can't be cut off by definition because the model is open-weighted," and that companies can deploy on-prem in their own ecosystem.

Token efficiency is the other lever for builders. Mistral reports matching GLM-5.3 on the Artificial Analysis Coding Agent Index with 47% fewer tokens per task and cutting Terminal-Bench output tokens by 30%, which it frames as making coding agents more economical to run at scale, with three reasoning modes exposed to trade performance against cost.

For security work, the Hugging Face blog notes that vulnerability research and incident response often look like attacks in model input, so closed source filters can block them, and that an API-dependent team can lose access mid-incident. Mistral says the model handles defensive tasks without unnecessary refusals and reports a 97.1% resistance rate on the B3 Agent Security Benchmark, stating that anyone hosting the model can run a safe agent without additional tuning. Licensing is a check to run first: the Mistral Ecosystem Open-Weight License permits commercial use, modification, and fine-tuning, but businesses operating a Model-as-a-Service offering must obtain a separate commercial license from Mistral.

What we don't know yet

The two sources give different numbers for the same test: Mistral's post states 51% on the Artificial Analysis Cyber Index, while the Hugging Face blog states 82% on an Artificial Analysis Cyber Index test that asks a model to reproduce a real vulnerability in open-source software and then patch it. Which configuration each figure refers to is unclear. The benchmark results, including Cognition's private FrontierCode 2.0 score and the Surge Code blind pairwise preferences, are reported by Mistral and not independently verified in the provided sources. Other unknowns: what the two official variants differ on, when the full technical report lands, what hardware is needed to self-host a 1T-parameter model, and what the specialized cybersecurity variant offered to design partners such as Orange Cyberdefense contains.

Sources: mistral.ai / TechCrunch AI / Hugging Face

How this Deep Take is made: it is written with AI assistance from the linked sources, then checked against them sentence by sentence, and every fact links to where it came from.