Fastino, led by Khosla, raised $17.5M to train AI models on cheap gaming graphics cards.
Tech giants love to boast about AI models with trillion-parameters that require massive GPU clusters. But Fastino has a different approach.
Palo Alto startup claims to have invented a new type of AI model architecture, which is small and task-specific. Fastino claims that the models are so small, they can be trained using low-end gaming graphics cards worth less than $100,000.
This method is attracting a lot of attention. Fastino, in an exclusive statement to TechCrunch, revealed that Fastino had secured $17.5M in seed funding, led by Khosla Ventures. Khosla Ventures was OpenAI’s famous first venture investor.
The startup now has a total funding of nearly $25 million. It raised $7m last November, in a preseed round led Microsoft’s VC arm M12 along with Insight Partners.
Ash Lewis, Fastino CEO and cofounder, says that “our models are faster, more precise, and cost a fraction of what flagship models do while outperforming them on specific tasks.”
Fastino sells a small model suite to enterprise customers. Each model is designed to address a specific need of a company, such as redacting sensitive information or summarizing documents. Fastino has not disclosed any early metrics or users, but its performance is astonishing early users. Lewis told TechCrunch that because the models are so small, they can deliver a complete response in just a few milliseconds.
Techcrunch event
Berkeley, CA|June 5
Is Fastino’s method going to catch on? It’s too early to tell. Companies like Cohere, Databricks and others are also promoting AI that excels in certain tasks. The enterprise-focused SATA models makers, such as Anthropic, Mistral and Mistral, offer small models. It’s no secret either that the future of generative AI in enterprise is likely to be smaller, more focused language model.
Time will tell, but a vote of confidence by Khosla is certainly helpful. Fastino is focusing on building an AI team that’s cutting-edge for the moment. It targets researchers at top AI laboratories who aren’t obsessed by building the biggest model, or beating benchmarks. Lewis says that “our hiring strategy is focused on researchers who have a contrarian thinking process about how language models are built right now.”
Charles Rollet, a senior reporter for TechCrunch. His investigative reporting led to U.S. sanctions against four tech firms, including China’s biggest AI firm. Charles covered the surveillance sector for IPVM before joining TechCrunch. Charles lives in San Francisco and enjoys hiking with dogs. You can reach Charles on Signal using charlesrollet.12, or by calling +1-628-282-2811.
View Bio
