Mistral Large 4 lands in preview, with open weights promised for the end of October

Mistral announced Mistral Large 4 on 6 October, a model it nicknames “le Chonk”. It carries one trillion parameters with 49 billion active per token, built as a hybrid instruct-and-reasoning mixture of experts, and it is natively multimodal. The model is in preview on Mistral’s API now, and Mistral says it will release the weights by the end of October.

The company’s claim is specific: Large 4 “significantly outperforms” any open-weight model developed in the US or Europe. Its published scores include 93% on Cybench, 61.7% on DeepSWE v1.1, 59.9% on AutomationBench and 28.3% on Terminal Bench 4.0, measured against named rivals including Claude Opus 5.5, DeepSeek V4 Pro, GLM-5.3, Kimi K3 and Qwen3.8 Max. On one visual-grounding benchmark it reports 42% against GPT-6-Astra’s 41%. heise reports the model was trained on 3,800 Nvidia Grace Blackwell GPUs.

The gap Mistral does not close is coding. heise’s assessment is that Large 4 is respectable against other open-weight models on software development and significantly behind the closed frontier models from OpenAI and Anthropic. Mistral’s own framing concedes the same point — the claim is best open-weight model, not best model.

For anyone picking a European model on sovereignty grounds, the weights release is the part that counts, because open weights are what make self-hosting on European infrastructure possible at all. Mistral has not stated which licence it will use. Aleph Alpha shipped Kolibri three days earlier, a 78-billion-parameter open-weight model aimed at organisations that want to run AI on their own hardware, so two of Europe’s frontier labs released openly inside a week.

The products a reader can use today are Vibe, Mistral’s assistant, and Vibe for code, its coding agent. API access to Large 4 is priced at $1.36 per million input tokens and $4.18 per million output. Mistral closed a roughly $3.4 billion Series D in September 2026 at a valuation near $24 billion.