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

Tech giants similar to boast astir trillion-parameter AI models that necessitate monolithic and costly GPU clusters. But Fastino is taking a antithetic approach.

The Palo Alto-based startup says it has invented a caller benignant of AI exemplary architecture that’s intentionally tiny and task-specific. The models are truthful tiny they’re trained with low-end gaming GPUs worthy little than $100,000 successful total, Fastino says.

The method is attracting attention. Fastino has secured $17.5 cardinal successful effect backing led by Khosla Ventures, famously OpenAI’s archetypal task investor, Fastino exclusively tells TechCrunch. 

This brings the startup’s full backing to astir $25 million. It raised $7 cardinal past November successful a pre-seed circular led by Microsoft’s VC limb M12 and Insight Partners.

“Our models are faster, much accurate, and outgo a fraction to bid portion outperforming flagship models connected circumstantial tasks,” says Ash Lewis, Fastino’s CEO and co-founder.

Fastino has built a suite of tiny models that it sells to endeavor customers. Each exemplary focuses connected a circumstantial task a institution mightiness need, similar redacting delicate information oregon summarizing firm documents.

Fastino isn’t disclosing aboriginal metrics oregon users yet, but says its show is wowing aboriginal users. For example, due to the fact that they’re truthful small, its models tin present an full effect successful a azygous token, Lewis told TechCrunch, showing disconnected the tech giving a elaborate reply astatine erstwhile successful milliseconds. 

It’s inactive a spot aboriginal to archer if Fastino’s attack volition drawback on. The endeavor AI abstraction is crowded, with companies similar Cohere and Databricks besides touting AI that excels astatine definite tasks. And the enterprise-focused SATA exemplary makers, including Anthropic and Mistral, besides connection tiny models. It’s besides nary concealed that the aboriginal of generative AI for endeavor is likely successful smaller, much focused connection models.

Time whitethorn tell, but an aboriginal ballot of assurance from Khosla surely doesn’t hurt. For now, Fastino says it’s focused connected gathering a cutting-edge AI team. It’s targeting researchers astatine apical AI labs who aren’t obsessed with gathering the biggest exemplary oregon beating the benchmarks.

“Our hiring strategy is precise overmuch focused connected researchers that possibly person a contrarian thought process to however connection models are being built close now,” Lewis says.