Ministral 3

Available · Language, Multimodal, Reasoning · Milestone

Ministral 3 is a series of dense Mistral AI language models in 3B, 8B and 14B sizes aimed at edge and local use, released on 2 December 2025 with Mistral Large 3. Each size has base, instruct and reasoning checkpoints with image input under Apache 2.0; Mistral's paper says all were derived from Mistral Small 3.1 by pruning and distillation. [1] [2] [3]December 2 (2025) Primary source

Timeline of Ministral 3 →

Claims and evidence

  • Released Primary source[1] [3]December 2 (2025)
  • Status Available Primary source[4]Generalist models [5]
  • Successor of Ministral EditorialEditorial link along the Ministral line: the announcement does not name a predecessor; the models overview lists Ministral 3 3B and 8B as the alternatives for the retired Ministral 3B and 8B of October 2024. Primary source[1] [4]Deprecated & retired models table, Alternative column
  • Derived from (distillation) Mistral Small 3.1The paper says all Ministral 3 models descend from Mistral Small 3.1 through Cascade Distillation: iterative pruning plus continued pretraining with logit distillation from the parent. The vision encoder is copied from Mistral Small 3.1 Base. Primary source[2]section 1 Introduction; Figure 1
  • Change · Size Adds a 14B size alongside 3B and 8B.Compared with Ministral Primary source[1]
  • Change · Reasoning Each size comes with a separate reasoning checkpoint next to the base and instruct checkpoints.Compared with Ministral Primary source[1]
  • Change · Modality All checkpoints accept image input through a 410M-parameter vision encoder copied from Mistral Small 3.1 Base.Compared with Ministral Primary source[2]
  • Change · Training data Trained by pruning and distilling from Mistral Small 3.1 on between 1 and 3 trillion tokens, according to the paper.Compared with Ministral Primary source[2]
  • Input text, image Primary source[6]Modalities [7]Modalities [8]Modalities [1]Ministral 3 section
  • Output text Primary source[6]Modalities [7]Modalities [8]Modalities
  • Feature Function calling Primary source[6]Features: Function Calling [7]Features: Function Calling [8]Features: Function Calling
  • Open weights YesBase, Instruct and Reasoning weights for all three sizes under Apache 2.0. Primary source[5] [1]
  • Context window 256k tokensThe paper gives context lengths up to 256k tokens, and 128k for the reasoning checkpoints. Primary source[6]Context [7]Context [8]Context [2]section 1 Introduction
  • Access API, Open-weights download, cloud partner Primary source[1]Available Today [5]

Lineage

Predecessors

Successors

No known successor.

Based on

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

Variants

Ministral 3 3B

Same dates as Ministral 3.

  • Variant Inline variant in this record.Base, Instruct and Reasoning checkpoints of this size are released; they are listed as aliases, not as separate variants. Primary source[6] [5]

Also known as: ministral-3b-2512, Ministral-3-3B-Instruct-2512, Ministral-3-3B-Base-2512, Ministral-3-3B-Reasoning-2512

Differs in:

  • Parameters: 3B Evidence not assessed [1] [2]Abstract

Ministral 3 8B

Same dates as Ministral 3.

  • Variant Inline variant in this record.Base, Instruct and Reasoning checkpoints of this size are released; they are listed as aliases, not as separate variants. Primary source[7] [5]

Also known as: ministral-8b-2512, Ministral-3-8B-Instruct-2512, Ministral-3-8B-Base-2512, Ministral-3-8B-Reasoning-2512

Differs in:

  • Parameters: 8B Evidence not assessed [1] [2]Abstract

Ministral 3 14B

Same dates as Ministral 3.

  • Variant Inline variant in this record.Base, Instruct and Reasoning checkpoints of this size are released; they are listed as aliases, not as separate variants. Primary source[8] [5]

Also known as: ministral-14b-2512, Ministral-3-14B-Instruct-2512, Ministral-3-14B-Base-2512, Ministral-3-14B-Reasoning-2512

Differs in:

  • Parameters: 14B Evidence not assessed [1] [2]Abstract

Related AI Radar coverage

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

All model releases from Mistral AI on AI Radar →