COUNE claims that the New Aya Vision AI model is the best class

CO for AIAI BOOKU BOOKU Non-profit research laboratory, this week claimed that a multimodal “Open” AI model, the Aya Vision, Laboratory was the best classroom.

The Moon Vision can answer the tasks such as writing tasks such as answering questions about pictures, to answer questions about pictures and create a summary of the main languages. Verdo-vision, which is free via WhatsApp, also called “a significant step to make technical progress for researchers around the world.”

“Although AI is an important progress, there is a large gap that models in different languages ​​- how many text and images are even more important in multimodal tasks,” Blog Post. “The moon is aimed at opening this gap to close this gap.”

The Moon meeting is in several flavors: Aya Vision 32b and Vision Vision 8b. The more the development of the two, the Moon has a “new border”, including the size of 2x, including the size of 2x, he said Meta’s Llama-3.2 90B Vision in certain visual understanding criteria. Meanwhile, according to the month meeting, according to Colare, according to the size of 3x, more evaluations of 3X measure scores better points.

Both models are available Embrace the face under a creative Commons 4.0 license from the AI ​​Dev platform COUNE’s acceptable use attachment. Cannot be used for trading applications.

COUNE, Moon vision, said that the English databases used to create translated and synthetic annotations are trained using the “different pool”. Annotes, also known as tags or labels, understand and interpret the models during the training process. For example, annotation to cultivate an image recognition model, can take a picture of the marks around the facilities or headings belonging to each person, ground or object.

COONE Using synthetic annotations – that is, the moments created by AI – are in Trend. In spite of the potential reductionRivals, including Openai, are increasingly preparing synthetic data It dries real world data. Research firm Gartner Approximate Last year, 60% of the information used for AI and analytical projects were created synthetically.

According to the region, the Moon meeting was a training on synthetic annotation, the laboratory allowed the laboratory to use the less resource.

“This is less of our critical attention to the less calculation and more computing using the less account,” he said. “It also supports more support, which allows you to calculate resources more often than the research community.”

Together with the Moon Vision, Coone, at the same time “vision-language”, a new benchmark suite designed to test the differences between screenshots and identify the differences between the two pictures.

AI Industry is in the middle of which some of the “evaluation crisis” Provide aggregate scores that make poor coordinating professional AI users are interested in tasks. Cohere, Ayavisionbench claims to be a step to provide a frame “wide and challenging” to evaluate the concept of a model and multimodal concept.

It is really the case with any chance.

“(T) DataSeti serves as a solid benchmark to assess vision languages ​​models in multilingual and real world parameters,” said researchers wrote in the post in hugging the face. “We make this assessment set that exists in this assessment community to put forward multilingual multimodal assessments.”