On October 6, 2026 Mistral put Mistral Large 4 into public preview — its largest model so far and, by Mistral’s own claim, the strongest open-weight model developed in the US or Europe. The weights are promised for the end of October; as of mid-October the model is available only through Mistral’s API.
Key facts
| Mistral Large 4 (preview) | |
|---|---|
| Parameters | 1.05T total, 52B active, 1.6B vision encoder |
| Context | 1M tokens |
| Input | Text and image |
| Price (in / out per 1M) | $1.36 / $4.18; 50% preview sale at $0.68 / $2.09 |
| Weights | Promised by end of October; license not yet published |
| Training | 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centers |
It is the first milestone on the roadmap Mistral funded with its €3 billion Series D in early September.
The benchmark picture
Mistral reports DeepSWE v1.1 61.7% and Terminal-Bench 4.0 28.3% — well below the best open and closed coding models — but leads on the work it is positioning for: a top-five spot on the Artificial Analysis Cyber Index, 59.9% on AutomationBench (ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro in Mistral’s table) and state of the art among open models on finance and law. In a blind Surge AI human evaluation of coding quality it ranked second of five (3.74), behind only Claude Opus 5 (4.22).
Why it matters
- A Western open-weight challenger at the 1T scale. Many of 2026’s largest open-weight releases — Kimi K3, Qwen3.8-Max, GLM-5.3, DeepSeek V4 — came from Chinese labs; Mistral is offering a sovereign alternative with European hosting.
- A cyber-first positioning. Mistral says its safeguards still allow legitimate security work — it scores 82% on a vulnerability reproduce-and-patch task where Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse. That is as much a statement about refusal policy as capability, and it is Mistral’s own claim.
- Pricing sits mid-pack. At $1.36 / $4.18 it is cheaper than Grok 4.7, Claude Sonnet 5.5 and GPT-6.1 Sol on list price, but well above DeepSeek V4.1-Flash and the Qwen3.8 Flash models.
Caveats
- It is a preview — no weights, license or full technical report yet, and preview prices are provisional.
- Almost every number is Mistral-reported, and the cyber comparison depends on how often closed models refuse.
- Coding scores trail the leading models on public benchmarks.
For the wider model landscape, see the 2026 LLM guide.