Tech giants often promote trillion-parameter AI models requiring significant investment in GPU clusters. However, Fastino is taking a different path. The startup, based in Palo Alto, claims to have developed a new AI model architecture that is intentionally small and task-specific. According to Fastino, these models can be trained using low-end gaming GPUs costing less than $100,000 in total.
This innovative approach is gaining attention. Fastino has reportedly secured $17.5 million in seed funding, led by Khosla Ventures, known as OpenAI’s first venture investor. This addition raises the company’s total funding to nearly $25 million, following a $7 million pre-seed round last November led by Microsoft’s VC arm M12 and Insight Partners.
Fastino’s CEO and co-founder, Ash Lewis, stated that these models are faster, more accurate, and cost-efficient, outperforming flagship models in specific tasks. The company offers a suite of small models for enterprise customers, focusing on specific needs such as redacting sensitive data or summarizing corporate documents.
While Fastino has not disclosed early metrics or users, it claims positive feedback from initial users. Due to their small size, the models can deliver entire responses in a single token, providing detailed answers almost instantly.
The enterprise AI sector is competitive, with companies like Cohere and Databricks also promoting task-specific AI models. Companies like Anthropic and Mistral provide small models as well. It is broadly anticipated that the future of generative AI in enterprises lies in smaller, focused language models.
While it remains to be seen if Fastino’s approach will become widely adopted, the endorsement from Khosla Ventures is a positive sign. Currently, Fastino is concentrating on building an advanced AI team, targeting researchers who may have unconventional ideas about language model development.
The company’s hiring strategy prioritizes researchers who think differently about constructing language models, according to Lewis.
