Mistral AI open models

vMistral Small 4 (26.03), Ministral 3 and Mistral Large 3 (25.12)Open sourceReleased 16 Mar 2026

Apache-licensed Mistral models, from the compact Ministral 3 up to Mistral Small 4. By Mistral AI SAS

Web app

Official sourceNo bundles

By the ToolsBharat teamUpdated 16 Mar 2026How we check

Quick answer

Mistral AI open models is an open source web app from Mistral AI SAS. Apache-licensed Mistral models, from the compact Ministral 3 up to Mistral Small 4. Download it only from huggingface.co. The latest version we track is Mistral Small 4 (26.03), Ministral 3 and Mistral Large 3 (25.12) (16 Mar 2026). ToolsBharat links straight to the publisher's own download page, with no bundled offers.

Latest version:: Mistral Small 4 (26.03), Ministral 3 and Mistral Large 3 (25.12) (16 Mar 2026)
Publisher:: Mistral AI SAS
Official site:: docs.mistral.ai
Licence:: Open source — Apache 2.0 (Mistral Small 4, Mistral Large 3, Ministral 3); Mistral Large 4 uses the separate Mistral Ecosystem Open Weight License
Requirements:: Ministral 3 8B fits in 12 GB of VRAM in FP8 (less if quantised further); Mistral Small 4 needs at least 4x NVIDIA HGX H100, 2x HGX H200 or 1x DGX B200

Our verdict

Mistral releases several of its models under Apache 2.0, and the Ministral 3 family is a practical pick for a single consumer GPU. Mistral Small 4 is far bigger than its name suggests and needs data-centre GPUs. Check each model's licence, because Mistral's newest flagship, Large 4, is not under Apache 2.0.

Key features

Ministral 3 in 3B, 8B and 14B sizes, each with Base, Instruct and Reasoning variants
Mistral Small 4: 119B mixture-of-experts with about 6B active parameters and configurable reasoning effort
256K-token context and image input on Ministral 3 and Small 4
Mistral Large 3 also under Apache 2.0

Pros

Pro: Apache 2.0 on the main open models
Pro: Ministral 3 8B fits in about 12 GB of VRAM in FP8, per Mistral

Cons

Con: Mistral Small 4 needs at least four H100s, two H200s or one DGX B200
Con: Licences differ between models; Large 4 is not Apache 2.0

What's new in Mistral Small 4 (26.03), Ministral 3 and Mistral Large 3 (25.12)

See what changed in Mistral Small 4 (26.03), Ministral 3 and Mistral Large 3 (25.12) in the official release notes

Version history

Version Mistral Small 4 (26.03), Ministral 3 and Mistral Large 3 (25.12)Latest
16 Mar 2026
SHA-256
The developer does not publish one
Source
Release notes

Requirements

OS: Ministral 3 8B fits in 12 GB of VRAM in FP8 (less if quantised further); Mistral Small 4 needs at least 4x NVIDIA HGX H100, 2x HGX H200 or 1x DGX B200
Architecture: Model weights (runs via a runtime such as Ollama, LM Studio or llama.cpp)

Licence

Type: Open source
Licence: Apache 2.0 (Mistral Small 4, Mistral Large 3, Ministral 3); Mistral Large 4 uses the separate Mistral Ecosystem Open Weight License
Publisher: docs.mistral.ai

Where these facts come from

Checked 30 Oct 2026

FAQ

How do I run a Mistral model locally?

For a single PC, start with Ministral 3. The Ollama library lists ministral-3 in 3B (3.0 GB), 8B (6.0 GB) and 14B (9.1 GB) tags. Mistral's model card for Mistral Small 4 also lists LM Studio as a supported way to run it.

User reviews

User reviewsयूज़र रिव्यू

No reviews yet

Used Mistral AI open models? Reviews appear after moderation; we never import or invent ratings.

Alternatives