Pixtral 12B

Available · Language, Multimodal · Milestone

Pixtral 12B is a multimodal model from Mistral AI that takes images and text and outputs text, combining a new 400M-parameter vision encoder with a 12B decoder based on Mistral Nemo. It was released under Apache 2.0 and offered on la Plateforme as pixtral-12b-2409 and in le Chat, and processes images at their native resolution within a 128k-token context window. [1] [2] [3]Modalities Primary source

Timeline of Pixtral 12B →

Claims and evidence

  • Released Other sources give: [3]release date Date on the docs model page. The blog post and the changelog give 2024-09-17 for the release on la Plateforme and le Chat. Primary source[1]page date; Use [4]September 17, 2024 [5]Eye of the Tiger - Pixtral on le Chat
  • Deprecated Primary source[6]Deprecated & retired models table, row Pixtral 12B [3]Deprecation date
  • Retired Primary source[6]Deprecated & retired models table, row Pixtral 12B
  • Status AvailableApache 2.0 weights remain downloadable from the mistralai Hugging Face organisation; the API endpoint pixtral-12b-2409 was deprecated on 2025-12-02 and retired on 2025-12-31. Primary source[7] [6]Deprecated & retired models table, row Pixtral 12B
  • Replaced by Ministral 3Mistral names Ministral 3 14B, a variant of the Ministral 3 record. Primary source[6]Deprecated & retired models table, row Pixtral 12B, Alternative [3]Replacement
  • Derived from (other) Mistral NeMoThe 12B multimodal decoder is built on Mistral Nemo 12B and paired with a new 400M vision encoder trained from scratch; the technical report says Mistral-Nemo 12B-Instruct initialised the decoder in its ablations. Primary source[1]Pixtral 12B in short: Architecture [2]section 2.1
  • Change · Modality Adds image input through a new 400M-parameter vision encoder trained from scratch; images are processed at their native resolution and aspect ratio.Compared with Mistral NeMo Primary source[1]
  • Change · Architecture Pairs that encoder with a 12B multimodal decoder built on Mistral Nemo; Mistral says it keeps text-only capabilities and can replace Mistral Nemo 12B as a drop-in.Compared with Mistral NeMo Primary source[1] [2]section 2.1
  • Input text, image Primary source[3]Modalities [1]Pixtral 12B in short
  • Output text Primary source[3]Modalities
  • Open weights YesApache 2.0. Primary source[1]Use; mistral-inference [7]
  • Context window 128K tokens Primary source[1]Architecture [3]Context
  • Parameters 12B12B multimodal decoder plus a 400M-parameter vision encoder. Primary source[1]Pixtral 12B in short: Architecture [7]Key features [3]Weights table
  • Access API, consumer app, Open-weights download Primary source[1]How to run Pixtral?

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 →