Fastino develops intentionally small, task-specific AI models that are trained using low-end gaming GPUs worth less than $100,000 in total. The company sells a suite of these models to enterprise customers for tasks such as redacting sensitive data and summarizing corporate documents. Fastino says its models are faster, more accurate on specific tasks, and far cheaper to train than flagship large models, with the ability to deliver an entire response in a single token in milliseconds. The startup has not disclosed early metrics or user counts but says early users are impressed with performance. Based in Palo Alto, Fastino is focused on hiring researchers at top AI labs who favor contrarian approaches to model design rather than trying to build the biggest models. The company recently raised $17.5 million in seed funding led by Khosla Ventures, bringing total funding to nearly $25 million. Fastino launched from stealth with a family of task-optimized language models engineered for enterprise use cases such as structuring textual data, RAG systems, text summarization, and task planning. The company says its architecture enables near-instant CPU inferencing and flexible deployment on CPUs or NPUs, reducing reliance on high-end GPUs and lowering energy consumption. Fastino claims its models are up to 1000x faster than traditional large language models and are designed to reduce hallucinations, adversarial vulnerability, and data-leakage risks. The product emphasis is on accuracy, speed, and safety for critical enterprise deployments. Fastino will use its announced funding to build an industry-leading research team, accelerate product development, and expand hiring from its Palo Alto headquarters. The company positions itself to help enterprises accelerate generative AI adoption by offering tunable, task-specific models.

Total raised

$25M

Funding rounds

2

Latest round

Seed

Latest activity

May 2025