AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuned / chat models

Magistral

Mistral AI
open weights / Overall score: 3.4

Mistral's reasoning model line, launched June 2025 in two tiers: the open-weight Magistral Small, a 24B model distilled from Magistral Medium traces with RL on top, and the API-only Magistral Medium. The current Small 1.2 adds a vision encoder and reasons in a dedicated thinking block, and fits on a single consumer GPU once quantized.

Mistral's reasoning line beside devstral (coding) and codestral (completion) on the map. Openness is read on the distributed SKU, Magistral Small, since Medium is API-only; the RL pipeline is described in the Magistral paper but not released.

Openness

3 high confidence
3.0
weights
open(Magistral Small 2506/2507/2509, ungated Apache-2.0 on HF)
post-training-data
described(SFT from Magistral Medium traces + RLVR, described in the paper, not released)
code
closed(no post-training pipeline released
license
Apache-2.0(OSI)

Every distributed Magistral Small release ships Apache-2.0 weights, ungated on the Hub, but the post-training mixture is only described in the paper, not released, and no pipeline ships. Magistral Medium, API-only, falls outside the distributed SKUs this reading covers, leaving the family open weights with a closed recipe.

  • https://huggingface.co/api/models/mistralai/Magistral-Small-2509 recorded 2026-10-07

    Model API for mistralai/Magistral-Small-2509: `gated: false`, `private: false`, cardData.license `apache-2.0`, base_model mistralai/Mistral-Small-3.2-24B-Instruct-2506, safetensors weights listed.

  • https://huggingface.co/mistralai/Magistral-Small-2509/raw/main/README.md recorded 2026-10-07

    Card: built on Mistral Small 3.2 (2506) "with added reasoning capabilities", "undergoing SFT from Magistral Medium traces and RL on top"; "Open license allowing usage and modification for both commercial and non-commercial purposes" under the Apache 2.0 heading; and a "Benchmark Results" table for Magistral Medium 1.2 and Small 1.2. No training mixture or pipeline is released.

  • https://mistral.ai/news/magistral recorded 2026-10-07

    The Magistral launch post announces a dual release: "Magistral Small — a 24B parameter open-source version and Magistral Medium — a more powerful, enterprise version", which scopes the license reading to the Small SKUs.

Adoption

2 high confidence
2.0

Adoption is measured as Hugging Face downloads summed across the three declared Magistral Small releases. Most of that volume still goes to the original 2506 release rather than the newer 2507 or 2509 updates.

Capability

4 medium confidence
4.0

Magistral Small posts AIME-class math results, 86.14% on AIME 2024, that place it level with DeepSeek-R1, a competitive open-weight reasoner short of the closed frontier. Magistral Medium, the tier's stronger member, scores higher still but is API-only and not counted since it is not distributed as weights.

Verified 2026-09-27