Mistral releases the open weights of Large 4, its trillion-parameter 'Le Chonk' model

Mistral AI has made the weights of Mistral Large 4, nicknamed "Le Chonk", available for download on Hugging Face. The French company had opened a public preview of the model through its API on October 6 and promised the weights for October 27; according to VentureBeat, they were published on October 29. The model is also offered through Mistral's own platforms, OpenRouter, NVIDIA-hosted endpoints and Microsoft Foundry.

What was released

Mistral describes Large 4 as a natively multimodal mixture-of-experts model that combines instruction-following and reasoning, with three selectable reasoning levels. Its documentation lists about 1.05 trillion parameters in total, 52 billion active, a 1.6-billion-parameter vision encoder and a context window of 1 million tokens. It takes text and images as input and returns text. Early reports of 49 billion active parameters reflected a different way of counting, VentureBeat notes.

The weights come under the Mistral Ecosystem Open-Weight licence, which the company presents as a modified MIT licence. It permits free use, modification, fine-tuning and commercial deployment, with one condition: businesses that offer the model as a hosted service must first obtain a separate commercial licence from Mistral. This is a change from Large 3, which was released under Apache 2.0.

Running the model locally is not light work. The Next Web reports that Mistral is also releasing a quantised version that runs on four Nvidia B200 or B300 GPUs; the company points to vLLM and SGLang for serving.

Benchmarks and pricing

Most performance figures come from Mistral or from the independent evaluator Artificial Analysis. Mistral reports 41% on Terminal-Bench 4.0, while VentureBeat says Artificial Analysis measured 25% on the same test. On the Artificial Analysis Coding Agent Index, the model scores 48, close to GLM-5.3, while using about 47% fewer tokens per task. Artificial Analysis also puts its cost per Intelligence Index task at $2.06, 63% below the preview, a figure that excludes self-hosting costs.

API prices have roughly halved compared with the preview, to $0.70 per million input tokens and $2.10 per million output tokens.

Open questions

Mistral says that after the preview it observed attempted real-world cyberattacks that the model refused, but also some potential for uplift, and that it is adjusting its guardrails. A specialised cybersecurity variant is available only to vetted partners. Independent results on several of the company's own benchmarks are still missing, and smaller distilled versions are promised without firm dates.