Mistral Large 4 weights released, matching top Chinese models

Mistral published the weights of Large 4 on Hugging Face on Thursday, October 29, letting anyone download the model and run it on their own infrastructure. A quantized version, compressed to cut memory needs, was released at the same time.

The preview presented on October 6 had fallen short in independent tests on coding and agent tasks. The final version closes much of that gap. On Terminal-Bench 4.0, which measures agentic coding, Large 4 climbs from 28% for the preview to 41%, according to charts shown to the press. GLM-5.3, the best Chinese open model on the chart, scores 40%, DeepSeek V4.1 Flash 31% and Kimi K3 21%. Mistral calls it the top open model for agentic coding on the Artificial Analysis index and the best open model built in the West.

Fewer tokens, a cyber pitch

Mistral says Large 4 uses 47% fewer tokens per task than GLM-5.3 on that index, and half as many as the preview on Terminal-Bench. Since AI is billed per token, that matters for cost. The company also added shorter "low" and "medium" reasoning modes alongside "high".

The model, nicknamed "Chonk", has 1,000 billion parameters, of which 52 billion are active per generated token. It accepts images but answers only in text and code. Mistral says reinforcement learning on internally generated coding and cybersecurity problems drove the gains, using 3,800 Nvidia Grace Blackwell GPUs in its European data centers.

On CyberGym-E2E, which asks a model to reproduce and fix a real vulnerability in open-source software, Large 4 scores 79%, level with Xiaomi's MiMo-V2.6-Pro. Claude Opus 5.5 and GPT-6 Astra score close to zero because they refuse the task, which partly reflects their safeguards. Mistral did not detail the safeguards in the released weights.

Still behind US closed models

The gap with American closed models remains wide: Opus 5.5 and GPT-6 Astra sit between 59% and 65% on Terminal-Bench, per measurements from early October. Google also raised the bar on Wednesday with Gemini 4 Carbon.

Running the quantized version, about 690 GB in 4-bit, requires four Nvidia B200 or B300 GPUs. Mistral promises a lighter version for individuals within a month. Pricing was not announced; the preview was listed at $1.36 per million input tokens and $4.18 per million output tokens. Large 4 is available on Hugging Face and through Mistral's API.