COHERE claims that the new Aya Vision AI model is the best in its class

Comrie AIAI Startup Coher’s Cohere’s non -profit, this week has released a multimedia “open” model, Aya Vision, the demanding laboratory is the best in its class.

Aya VISION can perform tasks such as writing photo comments, answering questions about images, translating text and generating summaries in 23 main languages. “An important step towards making technical inventions within the reach of researchers all over the world.”

“While artificial intelligence made great progress, there is still a big gap in the extent of models’ performance across different languages ​​- one becomes more clear in multimedia tasks that involve text and pictures.” Blog post. “Aya VISION aims to help in filling this gap.”

Aya Vision comes in a few flavors: Aya VISION 32B and Aya VISION 8B. Cohere said, more advanced than Monday, Aya Vision 32B Meta’s Llama-3.2 90B Vision On some criteria, visual understanding. Meanwhile, Aya VISION 8B is studying better in some 10x assessments of its size, according to Cohere.

Both models available From the AI ​​Dev Huging Face platform under Creative Commons 4.0 with Colrie Addented Adandum. It cannot be used for commercial applications.

Cohere said that Aya Vision was trained using a “variety” of English data collections, which the laboratory translated and used to create artificial comments. An illustrative comments, also known as marks or stickers, help models to understand and interpret data during the training process. For example, the explanatory suspension may take the form of image recognition form signs about organisms or illustrations that indicate every person, place or image of a photo in the image.

The use of artificial illustrations – that is, the explanatory comments created by artificial intelligence – in the direction. Despite the possible negative aspectsCompetitors, including Openai, are increasingly benefiting from artificial data to train models such as Well, realistic data dries up. Gartner research company Estimates 60 % of the data used for artificial intelligence projects and analyzes last year were industrially created.

According to COHERE, Aya enabled the laboratory artificial illustrations to use less resources with competitive performance.

“This shows our decisive focus on efficiency and (doing) more with less accounting,” Cote wrote on its blog. “This also provides greater support for the research community, who often enjoy a limited access to account resources.”

Along with Aya Vision, COHERE also released a new new suite, Ayavisionbench, designed to search the model skills in “Language” tasks such as determining the differences between two forms and converting screenshots to software instructions.

The artificial intelligence industry is in the midst of what some called a “evaluation crisis”, as a result of the generalization of the standards Give the total grades that are badly related In tasks most users of artificial intelligence are interested in. COHERE emphasizes that ayavisionbench is a step towards correcting it, which provides a “wide and difficult” framework for assessing understanding across languages ​​and multi -means of the model.

With any luck, this is already the case.

(T) Data collection is a strong standard for evaluating vision language models in multi -language and realistic settings. Book In embraced face. “We make this evaluation group available to the research community to pay multimedia assessments.”