AI Radar

Magistral Small

Available · Language, Reasoning

Magistral Small is the 24B-parameter open-weights variant of Magistral, which Mistral AI calls its first reasoning model, released under Apache 2.0 and also offered as an API model. Its model card says it builds on Mistral Small 3.1, with supervised fine-tuning on Magistral Medium reasoning traces followed by reinforcement learning. [1] [2] [3]June 10, 2025 Primary source

Timeline of Magistral Small →

Claims and evidence

  • Released Primary source[1]page date; Availability [3]June 10, 2025 [4]release date
  • Deprecated Primary source[5]Deprecated table, Magistral Small 1.0 (magistral-small-2506)
  • Retired Other sources give: [6]Model Lifecycle table, Retirement Date The Mistral Legal Center lifecycle table gives October 31, 2025 as retirement date, the date the docs models overview lists as deprecation date. Primary source[5]Deprecated table, Magistral Small 1.0 (magistral-small-2506) [4]Retirement date
  • Status AvailableApache 2.0 weights are still downloadable from Mistral AI's Hugging Face organisation; the API model magistral-small-2506 was retired on 2025-11-30. The Legal Center lifecycle table lists the model as Retired. Primary source[2] [1]Availability
  • Replaced by Mistral Small 4Replacement named for the API model when checked on 2026-10-01. Primary source[5]Deprecated table, Magistral Small 1.0, Alternative [4]Replacement
  • Variant of Magistral Medium Primary source[1]Highlights
  • Derived from (fine tune) Mistral Small 3.1Supervised fine-tuning on reasoning traces from Magistral Medium, followed by reinforcement learning. The model card names Mistral Small 3.1 (2503) as the base; the paper refers to it as Mistral Small 3 (24B). Primary source[2]first paragraph and base_model metadata [7]section Magistral Small: RL on top of reasoning SFT bootstrapping
  • Change · Reasoning Adds long reasoning chains before the answer, through fine-tuning on Magistral Medium traces and reinforcement learning on top of Mistral Small 3.1.Compared with Mistral Small 3.1 Primary source[2] [7]
  • Input text Primary source[4]Modalities
  • Output text Primary source[4]Modalities
  • Open weights Yes Primary source[1]Availability [2]
  • Context window 40k tokensThe Hugging Face card states a 128k window but says performance may degrade past 40k and recommends a 40k maximum. Primary source[4]Context
  • Parameters 24B Primary source[1]Highlights [2]first paragraph [4]Weights table, Parameters (B)
  • Access Open-weights download, API Primary source[1]Availability [3]June 10, 2025 [4]API id magistral-small-2506

Lineage

Predecessors

No known predecessor.

Successors

Based on

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Mistral AI on AI Radar →