Magistral Small 1.2

Available · Language, Multimodal, Reasoning

Magistral Small 1.2 is a 24B-parameter reasoning model from Mistral AI, built on Mistral Small 3.2 and released on 17 September 2025 on Mistral's API and as Apache 2.0 open weights. It accepts text and images; Mistral says a quantized version fits on a single RTX 4090 or a 32GB MacBook, and the weights stayed available after the API retirement in July 2026. [1] [2]September 17 (2025) [3] [4]Deprecated & retired models table, Magistral Small 1.2 Primary source

Timeline of Magistral Small 1.2 →

Claims and evidence

  • Released Other sources give: [3]page header date The docs model page gives September 18, 2025; the dated changelog entry, used as the value, gives September 17. Primary source[2]September 17 (2025)
  • Deprecated API deprecation; the open weights remain available. Primary source[4]Deprecated & retired models table, Magistral Small 1.2 [3]Deprecation date
  • Retired API retirement; the open weights remain available. Primary source[4]Deprecated & retired models table, Magistral Small 1.2
  • Status AvailableApache 2.0 weights remain downloadable from Mistral AI on Hugging Face. On Mistral's API the model was deprecated on 2026-04-30 and retired on 2026-07-31. Primary source[1] [4]Deprecated & retired models table, Magistral Small 1.2
  • Replaced by Mistral Small 4Replacement on Mistral's API. Primary source[4]Deprecated & retired models table, Alternative column [3]Replacement
  • Successor of Magistral Small 1.1 Primary source[1]Updates compared with Magistral Small 1.1
  • Derived from (fine tune) Mistral Small 3.2The model card says it builds on Mistral Small 3.2 (2506), with supervised fine-tuning on Magistral Medium traces followed by reinforcement learning. Primary source[1]introduction
  • Change · Modality Adds a vision encoder, so the model accepts image input and can reason over images as well as text.Compared with Magistral Small 1.1 Primary source[1]Updates compared with Magistral Small 1.1
  • Change · Reasoning Reasoning content is wrapped in new [THINK] and [/THINK] special tokens, and the reasoning prompt is given in the system prompt.Compared with Magistral Small 1.1 Primary source[1]Updates compared with Magistral Small 1.1
  • Change · Other Mistral reports better LaTeX and Markdown formatting, shorter answers to easy general prompts and fewer infinite generation loops.Compared with Magistral Small 1.1 Primary source[1]Updates compared with Magistral Small 1.1
  • Input text, image Primary source[3]Modalities [1]Updates compared with Magistral Small 1.1; Key Features
  • Output text Primary source[3]Modalities
  • Feature Function calling Primary source[3]Features: Function Calling
  • Open weights YesApache 2.0. Primary source[1] [3]Weights
  • Context window 128k tokensThe model card notes that performance might degrade past 40k tokens. Primary source[3]Context [1]Key Features: Context Window
  • Parameters 24B Primary source[1]introduction
  • Access API, Open-weights download Primary source[3]Features: API endpoints; Weights [1]

Lineage

Predecessors

Successors

No known successor.

Based on

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

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 →