Mistral
Mistral AI
Mistral is France's reference lab for open models. The family pairs Mistral Large 3 for general use (675 billion parameters, 41 billion active, mixture of experts), Mistral Small 4 and Ministral 3 for compact formats, Magistral for reasoning and Devstral for code. It stands as the sovereign European option, released as open weights, most often under the Apache 2.0 licence (Mistral Large 3, Mistral Small 4, Ministral 3, Magistral Small, Devstral Small 2); Mistral Medium 3.5 and Devstral 2 123B fall under a Mistral AI-specific licence (Hugging Face repositories re-read on 2 September 2026).
Key points: Sovereign FR and EU · Large 3, Small 4, Magistral, Devstral · Open weights · Mostly Apache 2.0.
The family
Mistral pairs Mistral Large 3 for general use (675 billion parameters, 41 billion active, mixture of experts), Mistral Small 4 and Ministral 3 for compact formats, Magistral for reasoning and Devstral for code.
Mistral Large 4, in preview since October 6, 2026
Mistral Large 4, nicknamed "le Chonk", is a 1.05-trillion-parameter mixture of experts with 49 billion active per token (52 billion with embeddings and output layer), multimodal on input. It is only available through the API as a preview, serving a 524,288-token context against one million announced. Weights, licence and architecture are expected at the end of October 2026, with the Hugging Face card showing October 31: this sheet will be re-read when they ship. Our detailed analysis cross-checks the figures and computes the hardware needed.
Strengths
As France's reference lab, Mistral stands as the sovereign European option, released as open weights, most often under the Apache 2.0 licence (Mistral Large 3, Mistral Small 4, Ministral 3, Magistral Small, Devstral Small 2); Mistral Medium 3.5 and Devstral 2 123B fall under a Mistral AI-specific licence (Hugging Face repositories re-read on 2 September 2026).
Use cases
The family covers general-purpose use, code, and reasoning, with a European regulatory and linguistic footing.
Deployment
Served by vLLM or llama.cpp depending on size, from GPU servers to workstations, on sovereign hosting.
All models on the platform are interchangeable through a single gateway: switching models is as simple as changing one line of configuration, with no code to rewrite. On-premises execution remains the priority.
Official sources
- mistralai, official Mistral AI models on Hugging Face: Mistral Large 3, Mistral Small 4, Magistral, Devstral
- Mistral Large 4, official documentation card of the preview: 1.05 T parameters, 1 M context announced
- mistralai/Mistral-Large-3-675B-Instruct-2512, official card: 675 billion parameters, 41 billion active, Apache 2.0 licence
- Artificial Analysis, model comparison index
Frequently asked questions
Is Mistral an open model?
As France's reference lab, Mistral stands as the sovereign European option, released as open weights, most often under the Apache 2.0 licence (Mistral Large 3, Mistral Small 4, Ministral 3, Magistral Small, Devstral Small 2); Mistral Medium 3.5 and Devstral 2 123B fall under a Mistral AI-specific licence (Hugging Face repositories re-read on 2 September 2026).
Why is Mistral absent from the sizing pages?
The sizing pages compute a memory footprint from a model's parameter count. No verified count is published for the version we recommend, so the site does not size Mistral rather than put forward an approximate figure that would under-size or over-size a machine.
For scale, the 7 open models the site does size occupy from 16 to 1,610 GB in NVFP4, runtime margin included. See the full sizing matrix, the format comparison and the engine and model combinations.
Mistral still runs on the same platforms. Sizing is then based on the characteristics of the chosen model, to be established case by case with the vendor.