Mistral releases open weights for 1T-parameter Large 4

French AI company Mistral has released the weights of Mistral Large 4, its flagship model nicknamed "Le Chonk", so businesses, governments and researchers can download it and run it on their own hardware. The model first appeared in a public preview earlier in October; Mistral says the released version is an improved one, trained further with reinforcement learning on coding and defensive security tasks.

What the model is

According to Hugging Face's launch post, Mistral Large 4 is a mixture-of-experts (MoE) model with roughly 1 trillion total parameters, of which about 49 billion are active for each token. It accepts text and images, outputs text, and supports a context window of up to 1 million tokens. The official weights come in FP8 and NVFP4 versions, and community GGUF builds are already available for llama.cpp-style tools. Mistral says it can be deployed with vLLM or SGLang, and NVIDIA offers it as a NIM container.

Running it locally is not a home-PC job: even with only 49 billion active parameters, all trillion parameters must be held in memory. Our guide to how much VRAM you need for local AI explains why model size, not just active parameters, decides the hardware.

Aimed at defenders

Mistral's main pitch is defensive cybersecurity: finding vulnerabilities in a codebase, reproducing them, and patching them without breaking the software. The company reports a 51% score on the Artificial Analysis Cyber Index, which tests that reproduce-and-patch workflow, and an average 96.6% refusal rate on malicious cyber requests, which it says is the highest it measured among open models. These are Mistral's own figures and benchmarks; independent testing will tell how well they hold up.

Mistral VP of science Pierre Stock told TechCrunch that open weights mean access "can't be cut off by definition", letting organisations run the model on-premises during an incident. He also acknowledged to Reuters that once weights are out, "it's impossible to contain access", which is the trade-off at the heart of the open-weight debate. We cover that debate in our explainer on what open-weight AI models are.

Licence and pricing

The model uses the Mistral Ecosystem Open-Weight licence, a modified MIT licence. It is free to use, modify, fine-tune and sell, with one condition: companies offering it as a hosted model-as-a-service need a separate commercial licence from Mistral. Through Mistral's API, Large 4 costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. It is also on OpenRouter and in Mistral's Vibe coding tool.

The release lands the same week Google introduced Gemini 4 Carbon, a closed cybersecurity-focused model available only to selected defenders, which makes the contrast between the two approaches unusually clear.