OpenModelsHQ

Mistral AI

Paris, France

Mistral AI is based in Paris, France. We track 7 open-weight models from it, and the newest is Voxtral-Mini-4B-Realtime-Arabic, released 8 Oct 2026.

Models tracked
7
LLM 3 · Vision 1 · Speech 3
Newest release
Voxtral-Mini-4B-Realtime-Arabic
8 Oct 2026
Country
France
Releases timeline
7 Mistral AI releases from 20 Sep 2023 to 8 Oct 2026. The largest is Mistral-7B-Instruct-v0.1 at 7.2B parameters. Dot area follows total parameters; a ring means the size is not published.
Releases timeline: data
ModelCategoryReleasedParameters
Voxtral-Mini-4B-Realtime-ArabicSpeech8 Oct 20264.4B
Mistral-Large-4-1T-A52BVision30 Sep 2026Not published
Voxtral-Mini-4B-Realtime-2602Speech21 Jan 20264.4B
Voxtral-4B-TTS-2603Speech17 Nov 2025Not published
Mistral-7B-Instruct-v0.2LLM11 Dec 20237.2B
Mistral-7B-Instruct-v0.1LLM27 Sep 20237.2B
Mistral-7B-v0.1LLM20 Sep 20237.2B

Origin

Arthur Mensch, Guillaume Lample and Timothée Lacroix met while studying at the École polytechnique. In April 2023 they established Mistral AI, a French company headquartered in Paris, according to Wikipedia. Mensch had worked at Google DeepMind; Lample and Lacroix had worked at Meta Platforms.

How it got here

Mistral 7B, the first model, was released on 27 September 2023 under Apache 2.0, which the announcement says means it can be used without restrictions. Mistral 7B Instruct v0.2 came next. Mixtral 8x7B, a sparse mixture-of-experts model with open weights, came on 11 December 2023, also under Apache 2.0.

Codestral broke the pattern. The 22B code model, announced on 29 May 2024, was released under a new Mistral AI Non-Production License that allows research and testing, with commercial licences available on request.

Magistral Small (open-source) and Magistral Medium, Mistral’s first reasoning models, came out on 10 June 2025. On 2 December 2025 the company released Mistral 3: Mistral Large 3, a mixture-of-experts model with 41B active and 675B total parameters, and the Ministral 3 series in 3B, 8B and 14B sizes. Wikipedia dates Mistral Small 4 to March 2026, under Apache 2.0.

There are speech models too. Voxtral Mini 4B Realtime is on Mistral’s Hugging Face page, and Wikipedia describes Voxtral TTS as an open-weights text-to-speech model with zero-shot voice cloning.

Licensing pattern

Apache 2.0 is Mistral’s main licence for open weights. Mistral 7B, Mixtral 8x7B, Mistral Large 3 and Ministral 3 all use it. The Mistral 3 announcement says both the base and instruction-tuned versions of Large 3 went out under it, and Wikipedia lists it for Mistral Small 4 as well.

Two exceptions stand out. Codestral carried the Non-Production License at launch. The Voxtral TTS card lists CC BY-NC 4.0, a non-commercial licence, and says the reference voices that come with the model use the same licence.

Mistral also runs hosted products. Wikipedia describes its Le Chat assistant, released in February 2024 and renamed Mistral Vibe in 2026. We list only the weights Mistral has published.

Release lineage

Newest first
  1. Voxtral-Mini-4B-Realtime-Arabic4.4B
  2. Mistral-Large-4-1T-A52B
  3. Voxtral-Mini-4B-Realtime-26024.4B
  4. Voxtral-4B-TTS-2603
  5. Mistral-7B-Instruct-v0.27.2B
  6. Mistral-7B-Instruct-v0.17.2B
  7. Mistral-7B-v0.17.2B

Largest first. The last line on each tile is the memory of the smallest device that fits the model at 4-bit.