Tech giants like to boast about trillions of parameter AI models that require large and expensive GPU clusters. But Fastino is taking a different approach.

The Palo Alto-based startup said it has invented a new AI model architecture that is intentionally small and task-specific. Fastino said the models were very small and they were trained on low-end gaming GPUs, with a total value of less than $100,000.

This method has attracted people's attention. Fastino exclusively told TechCrunch that Fastino has secured $17.5 million in seed funding led by Khosla Ventures, Openai’s first investor.

This brings the startup's total funding to nearly $25 million. It raised $7 million in a prepaid round led by Microsoft's VC ARM M12 and Insight Partners last November.

"Our models are faster and more accurate, and cost a small portion of training while outperforming flagship models on specific tasks," said Ash Lewis, CEO and co-founder of Fastino.

Fastino has built a small set of models that are sold to corporate customers. Each model focuses on specific tasks that a company may need, such as editing sensitive data or summarizing company documents.

Fastino has not disclosed early metrics or users, but says its performance surprised early users. For example, because their models are small, Lewis told TechCrunch that its model can provide the entire response in one token and demonstrates the technique in a detailed answer in milliseconds.

It is too early to tell if Fastino's method will continue to move forward. Enterprise AI space is crowded, and companies like Cohere and Databricks are also touting AI to stand out on certain tasks. Small models are also available from enterprise-centric SATA model manufacturers, including humans and Mistral. It is no secret that the future of enterprise-generated AI may be in smaller, more focused language models.

Time may tell, but Khosla's early vote of confidence certainly won't hurt. Currently, Fastino says it focuses on building a cutting-edge AI team. It is targeted at researchers at top AI labs who don't like to build the biggest models or beat benchmarks.

"Our recruitment strategy is very focused on researchers who may have a counter-trend thinking process, which is how to build language models now," Lewis said.