Pruna AI open sources its AI model optimization framework

Pruna has youA European Startup that is working on algorithms for AI models, creating its optimization structure Open source Thursday.

Pruna AI is creating a structure that applies a number of skills methods such as catching, trimming, quantization and plain, applied to the AI ​​model.

Pruna AI co-people and CTO John Raswan TechCrunch told TechCrunch, “We also apply to conservation and load of compressed models, apply combinations of this summary method, and even evaluate your compressed model even after compressing.”

Specifically, the structure of the Prun AI can evaluate if a model is after compressing and after compressing the profits you received, significant quality damage.

“If I use a metaphor how we hugged face standardized transformers and defizhers – how to call them, how to save them, how to save them, etc. We are doing the same, but for the sake of skill,” he added.

Large AI labs have already been using different summer methods. For example, the fastest version of Openai’s flagship models depends on the platform.

It probably developed a rapid version of OpenAI GPT -4 turbo, GPT -4. Likewise, the Flux 1-Fast Figure Generation Model is a subtile version of Flux 1 model from Black Forest Labs.

Distribution is a strategy used to collect knowledge from a large AI model with a “teacher-student” model. Developers send a request to a teacher model and records outputs. The answers are sometimes compared to a dataset to see how accurate they are. These outputs are then used to train the students to training the students, which are given approximate training for the teacher’s behavior.

“For big companies, what they usually do is that they make this thing at home and what you can find in the open source world is usually based on a single approach.

Pruna AI supports any type of model from any type of model to larger language model and any type of model in computer vision models, the company is now more specially focusing on the image and video generation models.

Pruna AI’s existing users include something Scene And PhotoreThe In addition to the Open Source version, there is an enterprise offer with advanced optimization features with an optimization agent of Pruna AI.

“We are expressing soon the most exciting feature will be an abbreviation agent,” said Raswan. “Basically, you give it your model, you said: ‘I want more speed but my accuracy is not over 2%’ ” And then, the agent will only find the best combination for you for it.

Pruna AI charges by the hour for its pro version. Raswan said, “When you rent a GPU on a GPU or a cloud service, how do you think about the GPU.”

And if your model is an important part of your AI infrastructure, you will save a lot of money to assume with the favorable model. For example, the pruna has created a Lama model eight times smaller without much damage using the AI’s contraction structure. Pruna AI hopes that its customers will think of its summary structure as an investment that pays for themselves.

Pruna AI raised a $ 6.5 million seed fund a few months ago. Startup investors include EQT Ventures, Duffney, Motiar Ventures and Kima Ventures.