Mistral releases the weights of Mistral Large 4

Mistral's news index dates the post 29 October 2026, and the page records a publication time of 29 October 2026, 12:00:07 UTC. The post says Mistral is releasing the weights of an improved Mistral Large 4, which it also calls Le Chonk. The post says the model was in public preview earlier in the month. It calls the model a 1-trillion-parameter, natively multimodal model. The Hugging Face card for mistralai/Mistral-Large-4-1T-A52B says it has 49 billion active parameters per token, or 52 billion including embeddings and output layers, a context length of 1 million tokens, and text and image input with text output. The card says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centres in Europe.

The post says the model is on the Mistral API, in Mistral Studio and on OpenRouter, and that the weights can be downloaded from Hugging Face and served with vLLM or SGLang. It also says the model is on NVIDIA-hosted endpoints and in an NVIDIA NIM container. The API name on the card is mistral-large-4. Starting with this release, the post prices the model at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. It says a full technical report is coming soon.

The post says the licence is the Mistral Ecosystem Open-Weight licence, which it describes as a modified MIT licence. It says the model is free to use, modify, fine-tune and commercialise, with one condition: a business that operates a model-as-a-service offering must obtain a separate commercial licence from Mistral before using the model commercially.

On benchmarks in the post, Mistral says the model scores 19.7% on Cognition's FrontierCode 2.0, 48% on the Artificial Analysis Coding Agent Index, which it describes as effectively tied with GLM-5.3 while using 47% fewer tokens per task on average, and 41% on Terminal-Bench 4.0, up 13 percentage points from the public preview, with output tokens reduced by 30%. It says the model scores 51% on the Artificial Analysis Cyber Index, and that its average refusal rate on malicious cyber requests is 96.6%, measured on the cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm. It says reasoning effort is low, medium, or high, and that high is the default.

The post says that, in live traffic since the preview, the model refused black-box attacks against live targets, and that Mistral sees a potential for uplift to conduct real attacks. It says it is modifying its guardrails and will keep updating them. It also says enterprise design partners, including Orange Cyberdefense, have preview access to a specialised variant for penetration testing, vulnerability assessment and red-teaming, with access subject to vetting and contractual safeguards, and with testing restricted to owned or explicitly authorised systems in controlled environments.

What it means for business: the weights are out, so a firm can run the model on its own infrastructure. The post's API prices are $0.70, $0.07 cached and $2.10 per million tokens. Offering the model as a service needs a separate commercial licence from Mistral. A variant aimed at penetration testing is limited to vetted partners and to systems they own or are authorised to test.