Tech giants often highlight trillion-parameter AI models that require large, costly GPU clusters; however, Fastino is adopting a different strategy.
The Palo Alto-based startup claims to have developed a novel AI model architecture that is deliberately small and task-specific. According to Fastino, these models are so compact that they can be trained using low-end gaming GPUs costing less than $100,000 in total.
This approach is gaining attention. Fastino has obtained $17.5 million in seed funding, led by Khosla Ventures, which was OpenAI’s initial venture investor, the company exclusively shared with TechCrunch.
This funding raises the startup’s total to nearly $25 million, following a $7 million pre-seed round from last November led by Microsoft’s VC arm M12 and Insight Partners.
“Our models are faster, more accurate, and cost-effective to train while outperforming flagship models on specific tasks,” stated Ash Lewis, Fastino’s CEO and co-founder.
Fastino has developed a series of small models marketed to enterprise customers. Each model targets a specific task a business might require, such as redacting sensitive data or summarizing corporate documents.
Although Fastino has not disclosed early performance metrics or user details, it reports impressive feedback from initial users. Due to their size, Fastino claims its models can provide comprehensive responses in a single token, delivering detailed answers in milliseconds, as explained by Lewis to TechCrunch.
It remains uncertain if Fastino’s approach will prevail. The enterprise AI market is competitive, with companies like Cohere and Databricks also offering task-specific AI solutions. Other small model producers, such as Anthropic and Mistral, are part of this space as well. The future of generative AI for enterprises is likely in smaller, more specialized language models.
Only time will determine the success of Fastino’s method. However, early support from Khosla is a positive sign. Currently, Fastino is concentrating on assembling a leading AI team, targeting researchers from top AI labs who are not necessarily focused on creating the largest models or exceeding benchmarks.
“Our hiring strategy is aimed at researchers with a contrarian perspective on current language model development,” Lewis added.
