Pruna AI open sources its AI model optimization framework

You haveA European startup, which works on compression algorithms for AI models, corrects the optimization frame open source Thursday.

ProNa creates several effectiveness methods such as AI, cache, pruning, quantity and distillation, a framework that applies to a certain AI model.

“We also appreciate the maintenance of compressed models, compiling these compromising methods, compressing the compatrible model, compressing the compressed model,” he said.

In particular, the Frame of the Pruna EU can evaluate if there is a significant damage after compressing a model and the performance you get.

“If I used a metaphor, how to get rid of the standard transformers and diffusers, how to get rid of them, how to save them, save them, etc.

Big AI laboratories have already used various compression methods. For example, Openai has become distillation to create faster versions of flagship models.

It is likely that Openai GPT-4 Turbo, a faster version of GPT-4. Similarly, Flux.1-fast Image generation model is the distilled version of the Flux.1 model from black forest laboratories.

Dedicate is a technique used to remove knowledge from a large AI model with a teacher-student model. The developers send a request to a teacher model and record their performances. Answers are sometimes compared to a database to see how true they are. These speeches are then used to train a trained student model to approximate the teacher’s behavior.

“For big companies, in general, it is what they are doing in the house.” But you won’t find a tool that connects them all, use and combine them together. This is a great value that Pruna brings right now. “

ProLa AI, from any model, supporting models from extensive language models to diffusion models, speech and textual models and computer imaging models, is in the center of more special photos and video generation models.

Some Pruna AI includes existing users Scenario and Photoroom. In addition to the open source issue, the Pruna AI presents an enterprise with advanced optimization features, including an optimization agent.

“The most interesting feature we have left soon will be a compression agent,” Reywvan said. “Basically, you give your model, you say: ‘I want more speed but don’t leave more than 2% of my accuracy.’ And then the agent will simply do the magic. It will find the best combination for you, it will return it to you. You don’t have to do anything like a developer. “

ProNa AI is charged for the Pro version for hours. “When you rent a GPU in AWS or any cloud service, it looks like you will think of a GPU,” Reywvan said.

If your model is a critical part of your AI infrastructure, you will end up with a lot of money with an optimized model. For example, Pruna AI, using a compression framework, made an eight-time zip model without much loss. Pruna AI hopes to think about his compression framework as an investment who pays for its customers.

ProNa AI won a dollar of $ 6.5 million for a few months ago. Initial investors include EQT Ventues, Daphni, Motier Ventures and whom businesses.