Fastino trains AI models on cheap gaming GPUs and just raised $17.5M led by Khosla

Technical giants like to boast about trillion parameters of AI models that require mass and expensive GPU groups. But Fastino makes a different approach.

Palo Alto-based start, deliberately small and task-specific AI model architecture says a new type. Fastino says models are as small as the low-level GPUs of GPU, which is less than $ 100,000 on holiday.

The method draws attention. Fastino provides $ 17.5 million to the seed funds led by Hosla enterprises, Openai’s first enterprise investor, Fastino Techcrunch.

This delivers the total financial financing for about $ 25 million. In November, $ 7 million, Microsoft trained M12 and Insight partners in a seeded tour.

“Our models are faster, more accurate and ahead of a fraketical models on special tasks on special tasks,” Ash Lewis, Fastinsin CEO and co-founder.

Fastino set up a set of small models sold to enterprise customers. Each model focuses on a certain task that a company needs a company such as reinstatement of sensitive data or generalization of corporate documents.

Fastino is still not yet disclosed to early size or users, but the early users of his performance are destroyed. For example, because they are so small, the models can answer a token in a Token, the Lewis TechCrunch can respond to the TechCrunch, once in a technology that once answered the detailed technology in Milice.

It’s still a little early if Fastino knows that he will approach. An entity is crowded as the worries of AI, which are superior to certain positions, such as AI space, Coone and Databricks. And also offers small models, including SATA model manufacturers, anthropical and autral. There is no secret of the future of a generative AI for the enterprise Probably in smaller, more careful language models.

Time can say, but the early confidence voting from Hosla does not hurt. Still, Fastino says it is aimed at building a leading AI team. Students are targeting in the best AI laboratories that are not obsessed by building the largest model or beat the criteria.

“Our hiring strategy is very directed to the researchers who have a contrabile thought process, which are probably how language models are built,” Lewis said.