Fastino trains AI fashions on low-cost gaming GPUs and simply raised $17.5M led by Khosla

Tech giants prefer to boast about trillion-parameter AI fashions that require large and costly GPU clusters. However Fastino is taking a special method.

The Palo Alto-based startup says it has invented a brand new type of AI mannequin structure that’s deliberately small and task-specific. The fashions are so small they’re skilled with low-end gaming GPUs value lower than $100,000 in whole, Fastino says.

The strategy is attracting consideration. Fastino has secured $17.5 million in seed funding led by Khosla Ventures, famously OpenAI’s first enterprise investor, Fastino solely tells TechCrunch. 

This brings the startup’s whole funding to almost $25 million. It raised $7 million final November in a pre-seed spherical led by Microsoft’s VC arm M12 and Perception Companions.

“Our fashions are sooner, extra correct, and price a fraction to coach whereas outperforming flagship fashions on particular duties,” says Ash Lewis, Fastino’s CEO and co-founder.

Fastino has constructed a collection of small fashions that it sells to enterprise prospects. Every mannequin focuses on a particular activity an organization may want, like redacting delicate knowledge or summarizing company paperwork.

Fastino isn’t disclosing early metrics or customers but, however says its efficiency is wowing early customers. For instance, as a result of they’re so small, its fashions can ship a complete response in a single token, Lewis informed TechCrunch, exhibiting off the tech giving an in depth reply directly in milliseconds. 

It’s nonetheless a bit early to inform if Fastino’s method will catch on. The enterprise AI area is crowded, with corporations like Cohere and Databricks additionally touting AI that excels at sure duties. And the enterprise-focused SATA mannequin makers, together with Anthropic and Mistral, additionally supply small fashions. It’s additionally no secret that the way forward for generative AI for enterprise is doubtless in smaller, extra targeted language fashions.

Time might inform, however an early vote of confidence from Khosla actually doesn’t harm. For now, Fastino says it’s targeted on constructing a cutting-edge AI staff. It’s concentrating on researchers at high AI labs who aren’t obsessive about constructing the largest mannequin or beating the benchmarks.

“Our hiring technique could be very a lot targeted on researchers that possibly have a contrarian thought course of to how language fashions are being constructed proper now,” Lewis says.