AI Radar

Aya Vision

Available · Language, Multimodal

Aya Vision is a multilingual vision-language model from Cohere For AI (now Cohere Labs) that takes text and images and produces text in 23 languages, released in 8B and 32B sizes as open weights, on the Cohere platform and through WhatsApp. The 32B model is built on Aya Expanse 32B, the 8B model on Command R7B. [1] [2]Model Details: Model Architecture [3]Model Details: Model Architecture Primary source

Timeline of Aya Vision →

Claims and evidence

  • Released Both pages show 4 March 2025; the blog's datePublished metadata is 2025-03-03 20:43 US Eastern time. Primary source[1]date line; closing availability paragraph [4]March 4, 2025
  • Aya Vision 8B retired from the Cohere API Primary source[5]2026-04-04: Embed v2.0, Aya Expanse 8B [6]
  • Status AvailableThe 32B model is Live on the Cohere API; the 8B model was retired from the API on 2026-04-04, but the weights of both sizes remain downloadable from Cohere Labs on Hugging Face. Primary source[7]Aya table: c4ai-aya-vision-32b, Live [2] [3]
  • Derived from (other) Aya ExpanseApplies to Aya Vision 32B: its language model is Aya Expanse 32B, paired with a vision encoder through a multimodal adapter. Primary source[3]Model Details: Model Architecture
  • Derived from (other) Command R7BApplies to Aya Vision 8B: its language model is based on Command R7B, further post-trained with the Aya Expanse recipe and paired with a vision encoder through a multimodal adapter. Primary source[2]Model Details: Model Architecture
  • Change · Modality Adds image input: Aya Vision accepts text and images, while Aya Expanse takes text only.Compared with Aya Expanse Primary source[3]Model Details: Input [8]Model Details: Input
  • Input text, image Primary source[2]Model Details: Input [3]Model Details: Input [7]Aya table: Modality Text, Images
  • Output text Primary source[2]Model Details: Output [3]Model Details: Output
  • Feature Multilingual23 languages. Primary source[1]second paragraph [3]Model Details: Languages covered
  • Open weights YesReleased under CC-BY-NC with Cohere Labs' Acceptable Use Policy; download is gated behind accepting the conditions. Primary source[1]Open-weights and research access [2] [3]
  • Context window 16K tokens Primary source[2]Context length [3]Context length [7]Aya table: c4ai-aya-vision-32b and c4ai-aya-vision-8b
  • Access API, Open-weights download, consumer appCohere platform, Kaggle and Hugging Face at release, plus free access through WhatsApp. Primary source[1]Open-weights and research access; closing availability paragraph

Lineage

Predecessors

No known predecessor.

Successors

No known successor.

Based on

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

Variants

Aya Vision 8B

Same dates as Aya Vision.

  • Variant Inline variant in this record. Primary source[2] [1] [9]Abstract

Status: Available Evidence not assessed [2] [5]2026-04-04: Embed v2.0, Aya Expanse 8B

Also known as: c4ai-aya-vision-8b, aya-vision-8b

Differs in:

  • Parameters: 8 billion parameters Evidence not assessed [2]Model Size [1]Open-weights and research access

Aya Vision 32B

Same dates as Aya Vision.

  • Variant Inline variant in this record. Primary source[3] [1] [9]Abstract

Status: Available Evidence not assessed [7]Aya table: c4ai-aya-vision-32b, Live [3]

Also known as: c4ai-aya-vision-32b, aya-vision-32b

Differs in:

  • Parameters: 32 billion parameters Evidence not assessed [3]Model Size [1]Open-weights and research access

Related AI Radar coverage

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

All model releases from Cohere on AI Radar →