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

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

The Palo Alto-based startup says it’s invented a brand new form 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 tactic is attracting consideration. Fastino has secured $17.5 million in seed funding led by Khosla Ventures, famously OpenAI’s first enterprise investor, Fastino completely 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 quicker, 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 clients. Every mannequin focuses on a selected 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 advised TechCrunch, displaying off the tech giving an in depth reply directly in milliseconds. 

It’s nonetheless a bit early to inform if Fastino’s strategy will catch on. The enterprise AI house is crowded, with firms 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 likely in smaller, more focused language models.

Time might inform, however an early vote of confidence from Khosla definitely doesn’t damage. For now, Fastino says it’s targeted on constructing a cutting-edge AI crew. It’s focusing on researchers at prime AI labs who aren’t obsessive about constructing the most important mannequin or beating the benchmarks.

“Our hiring technique may 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.