# Cohere released Aya Vision as an open-weight multilingual vision-language family Cohere Labs released Aya Vision on March 4, 2025, pairing 8B and 32B open-weight vision-language models with evaluation across 23 languages. Canonical: https://rangoon.ai/insights/archive-cohere-aya-vision-multilingual-march-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-24 Updated: 2026-10-24 Event date: 2025-03-04 Tags: [Cohere](https://rangoon.ai/insights/tags/cohere/), [Aya Vision](https://rangoon.ai/insights/tags/aya-vision/), [multilingual AI](https://rangoon.ai/insights/tags/multilingual-ai/), [multimodal AI](https://rangoon.ai/insights/tags/multimodal-ai/) Source note: Historical reporting: Cohere Labs announced Aya Vision on March 4, 2025 as an open-weight multilingual and multimodal model family. The release named 8B and 32B variants. A technical companion explains its 23-language evaluation, synthetic annotations, translation and model-merging methods. Cohere-reported benchmark comparisons are supplier measurements, not independent validation for another deployment. Image: Multilingual image request enters a selected Aya model and language-specific evaluation. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - Aya Vision made multilingual image-and-text evaluation part of the model release. - Open weights are distinct from a claim that every training input or deployment route is open. - Vendor benchmark results need their stated language set and method before use in a product decision. ## Cohere announced Aya Vision on March 4, 2025 [source 1](https://docs.cohere.com/changelog/aya-vision-announcement) Cohere Labs announced Aya Vision on March 4, 2025 as a multilingual, multimodal model family. The dated release identifies image captioning, visual question answering, text generation, and translation from text or images into text as intended tasks. It is a model release, not a statement that a product automatically supports every language or image workflow. The event record matters because it fixes the 8B and 32B family at a specific release point. It does not prove a particular application selected the correct model, prepared its images consistently, or met a target language’s quality threshold. ## The release paired open weights with a stated 23-language evaluation scope [source 1](https://docs.cohere.com/changelog/aya-vision-announcement) [source 2](https://huggingface.co/blog/aya-vision) The technical account says Aya Vision was released as open weights in 8B and 32B forms and evaluated across 23 languages. That pairing is useful but should not be collapsed into a single openness label. Weights permit a form of inspection and local adaptation; an evaluation language set identifies where the publisher measured behavior. Neither fact lists an adopter’s fine-tuning corpus, inference provider, or end-user language mix. A multilingual deployment therefore needs a language-by-task test matrix, including text-only, image-plus-text, and image-originated translation cases. A score from a general multilingual suite cannot stand in for a policy, legal, or domain-specific acceptance test. ## The companion description identifies data transformation and model merging as release methods [source 2](https://huggingface.co/blog/aya-vision) Cohere’s technical companion describes synthetic annotations, translation and rephrasing to expand multilingual data, along with multimodal model merging. Those are method claims that help distinguish Aya Vision from a simple text-model-plus-image-adapter description. They also explain why an observed answer should be tested as a joint language-and-vision outcome. For example, an image caption may be fluent in a requested language while still lose a visual qualifier, and a translation may preserve words while changing an instruction’s scope. The source describes the publisher’s approach; it does not supply a customer’s preprocessing settings, prompt template, safety filters, or audit record. Those belong to a deployment evaluation. ## Published comparisons are useful only with their benchmark and language conditions attached [source 1](https://docs.cohere.com/changelog/aya-vision-announcement) [source 2](https://huggingface.co/blog/aya-vision) Cohere reports pairwise results on AyaVisionBench and mWildVision, including comparisons with larger named vision models. The comparison is newsworthy because the release made multilingual benchmark design part of its product argument. It is still a vendor result under the publisher’s stated datasets, translations, prompts, and evaluation procedure. It does not establish that Aya Vision will outperform another model on a reader’s image source, language variant, or response format. A useful release review records model size, model revision, language, input resolution, benchmark version, and whether a result came from a human preference comparison or an automated metric. That makes a later local evaluation comparable instead of treating a headline rank as a universal capability claim. ## Sources 1. [Cohere: Aya Vision announcement](https://docs.cohere.com/changelog/aya-vision-announcement) 2. [Cohere For AI: Aya Vision technical deep dive](https://huggingface.co/blog/aya-vision) ## Related reading - [Cohere’s Command A Vision added image inputs with explicit request and tool-use limits](https://rangoon.ai/insights/archive-cohere-command-a-vision-july-2025/) - [Aya Expanse made multilingual evaluation a release requirement, not a translation claim](https://rangoon.ai/insights/archive-cohere-aya-expanse-multilingual-2024/) - [Tencent published the Hunyuan-MT technical report for multilingual translation models](https://rangoon.ai/insights/archive-tencent-hunyuan-mt-technical-report-september-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)