Prona you haveA European startup that is working on compression algorithm for AI models, making its correction framework Open source Thursday
Prona AI is creating a framework that applies a number of performance methods of performance such as catches, harvesting, quantization and oson on a given AI model.
“We also standardize the savings and loading of compressed models, apply the combination of these compression methods, and even after compressing your compressed model.”
In particular, Prona AI’s framework can estimate that if you have a significant loss of quality after compressing a model and getting the benefits of performance.
He added, “If I use a metaphor, we are similar to how to face the standard transformers and joints that are embraced – how to call them, how to save them, save them, how to save them, etc., etc. We are doing the same, but for performance.”
Big AI lab is already using various compression methods. For example, Openi has been relied on to make the fastest version of his flagship model.
It is likely that Openi developed GPT4 Turbo, a sharp version of GPT4. Similarly, Flux 1-Fast Image Generation Model is an east version of Black Forest Labs Floox 1 model.
Oson is a technique used to extract knowledge from a large AI model with a “teacher’s student” model. Developers send applications to the teachers’ model and record the results. The answers are sometimes compared with the data to see how accurate they are. These results are then used to train the student’s model, which is trained to estimate the teacher’s behavior.
Rachwan said, “What they usually do for big companies, is that they build this equipment at home. And in the world of open source, what you can find is usually based on the single methods. For example, a quantization method for LLM, or a method of disagreement models.” “But you didn’t find a device that collects them all, making it easy to use and connecting them all together. And this is the great value that the older is bringing right now.”
Although Prona AI supports any kind of model, from a large language model, from speech to text model and computer vision model, the company is still focusing more on image and video generation models.
Prana AI includes some current users Perspective And Photoorome. In addition to the open source edition, Prona AI has an enterprise offering that has modern correction properties, including a correction agent.
“The most interesting feature we are releasing soon will be a compression agent,” said Rachwan. “Basically, you give it your model, you say: ‘I want more speed but don’t leave my accuracy more than 2 %.’ And then, the agent will just do his magic, you don’t have to do anything as a developer.
Prona AI receives an hour for its supporter version. “When you rent GPU on AWS or a cloud service, how do you think about GPU, Rachwan said.
And if your model is an important part of your AI infrastructure, you will have a lot of money saving with a better model. For example, Prona AI has created an eight -times smaller lama model without much damage using her compression framework. Prona Ai hopes that her Grah will think of her compression framework as an investment that he pays for himself.
Prona Ai collected $ 6.5 million worth of seeds a few months ago. Startup investors include Act Ventures, Defani, Motor Ventures and Kima Ventures.