What happened
Mistral AI announced Mistral Large 4 on October 6, 2026, and opened a public preview through its API in Mistral Studio. The company describes it as an open-weight, natively multimodal, general-purpose model. The weights are not yet downloadable: Mistral says they will be released by the end of October 2026, together with more detail on architecture, benchmarks and post-training.
What’s new
- Architecture: granular mixture-of-experts with about 1.05 trillion total parameters and 52 billion active per token, plus a 1.6B-parameter vision encoder.
- Context window: 1 million tokens.
- Features: structured outputs, function calling, document Q&A, batching, agents and conversations endpoints, and built-in tools.
- Pricing (per million tokens): $1.36 input, $0.14 cached input, $4.18 output. The documentation page also shows a discounted rate of $0.68 input and $2.09 output, so check the console for the price applied to your account.
- Training: 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centers. Mistral will offer it in multiple regions, including a European deployment it runs end to end.
- License: not yet named. Mistral’s pages say only “open-weight”; its previous Large 3 model used Apache 2.0.
Mistral-reported results include 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4.0, 93% on Cybench and 59.9% on AutomationBench. These are vendor figures (some attributed to third-party evaluators) and have not yet been independently reproduced.
Why it matters
If the weights ship as promised, Large 4 would be one of the largest open-weight models available, and the only one in that class from a European lab with an EU-operated hosted option. That is relevant for organizations with data-residency or sovereignty requirements that want frontier-scale capability without sending data to US providers.
Two open questions affect practical adoption. First, the license: a permissive license would allow broad commercial self-hosting, while a custom license could restrict it. Second, hardware: a model with roughly a trillion total parameters needs multi-GPU server infrastructure even with only 52B active, so most teams will use it through an API or a cloud host rather than locally.
How to try it
- Sign in to Mistral Studio (console.mistral.ai) and select Mistral Large 4 in the playground.
- Call it via /v1/chat/completions using the model identifier shown in the docs (listed as mistral-large-4-0).
- Watch Mistral’s Hugging Face organization and the license file when the weights are published.