According to the Hugging Face blog, it is a Mixture-of-Experts model with about one trillion total parameters, roughly 49 billion active per token, and a one-million-token context window. It takes text and images as input and outputs text, and ships in two official variants, FP8 and NVFP4, alongside GGUF versions from Unsloth and the llama.cpp team. The license is the Mistral Ecosystem Open-Weight License, a modified MIT license: the model may be used, modified, fine-tuned and commercialized, but businesses offering it as a Model-as-a-Service need a separate commercial license from Mistral. Mistral says the model was trained from scratch on European infrastructure.
Mistral says it kept training the model with reinforcement learning in the three weeks since the preview. According to the company, it rose 13 percentage points to 41% on Terminal-Bench 4.0, leading open-weight models while using about 30% fewer output tokens. On the Artificial Analysis Coding Agent Index it scored 48%, effectively tied with GLM-5.3, using 47% fewer tokens per task on average. Three reasoning levels (low, medium and high) were added to balance performance against cost.
On cybersecurity, Mistral reports 51% on the Artificial Analysis Cyber Index, which tests finding, reproducing and patching vulnerabilities in code, and an average refusal rate of 96.6% on malicious cyber requests. The company adds that monitoring live traffic since the preview showed potential uplift for real attacks, so it is changing its guardrails. A specialized variant for penetration testing and red-teaming is available in early access to partners such as Orange Cyberdefense, after vetting and under contract.
The model is available in Mistral Studio, through the Mistral API and on OpenRouter, and as downloadable weights on Hugging Face for self-hosting with vLLM or SGLang. API pricing is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The figures come from the company and from Hugging Face and are not all consistent: Hugging Face cites 82% on a vulnerability reproduction-and-patching test, while Mistral cites 51% on the overall cyber index. The company did not say when the specialized cyber variant will open to a wider audience.
