Fastino trains AI models at cheap games and raised $ 17.5 million led by Khosla

Technological giants like to boast of the Billion Parameters models that require massive and expensive GPU groups. But Fastino It is adopting a different approach.

The Startup based in Palo Alto says that it has invented a new type of AI model architecture that is intentionally small and specific to tasks. The models are so small, they are trained with low -end games for a value of less than $ 100,000 in total, says Fastino.

The method is attracting attention. Fastino has obtained $ 17.5 million in initial funds led by Khosla Ventures, Openai’s first risk investor, Fastino tells Techcrunch exclusively.

This carries the total financing of the startup to almost $ 25 million. He raised $ 7 million last November in a round prior to the seed led by VC ARM M12 of Microsoft and Insight Partners.

“Our models are faster, more precise and cost a fraction to train while they exceed flagship models in specific tasks,” says Ash Lewis, CEO and co -founder of Fastinino.

Fastino has built a set of small models that sells to business clients. Each model focuses on a specific task that a company may need, such as writing confidential data or summarizing corporate documents.

Fastino does not reveal the first metrics or users, but says that his performance is surprising the first users. For example, because they are so small, their models can deliver a complete response in a single file, Lewis told TechCrunch, showing technology giving a detailed response at the same time in milliseconds.

It is still a bit early to know if Fastino’s approach will realize. The Enterprise space is full, with companies such as Cohere and Databricks also promoting the standing in certain tasks. And company -centered SATA model manufacturers, including anthropics and mystral, also offer small models. Nor is it any secret that the future of the generative AI for the company is Probably in smaller and more focused language models.

Time can say, but a vote of Khosla’s early trust certainly does not hurt. For now, Fastino says he focuses on building a vanguard team. It is aimed at Top AI Labs researchers who are not obsessed with building the largest model or overcoming the reference points.

“Our hiring strategy focuses a lot on researchers who may have a process of opposite thinking about how language models are being built at this time,” says Lewis.