Most AI voice agents still sound like they’re buffering their personality. Smallest.ai, a startup founded in late 2024, just raised a $13 million Series A to fix that — not by speeding up large language models, but by shrinking them. The round was led by Seligman Ventures, with Sierra Ventures and 3one4 Capital chipping in, pushing total funding past $21 million.
The core bet from founder and CEO Sudarshan Kamath is deliberately counterintuitive. While the industry races toward ever-larger models, Smallest.ai builds a compact voice model that processes speech the way humans do: listening, thinking, and speaking at the same time. “While I’m speaking to you, you’re already thinking, and you might interrupt me if I talk for too long,” Kamath explained. That’s the interaction the startup is engineering — a model that doesn’t politely wait for you to finish a complete audio clip before formulating a response.
When the small model hits the edge of its knowledge, it doesn’t hallucinate. It hands off to a larger foundational model, putting the customer on a brief hold to “research” — a design choice that mirrors what a human agent would actually do. Kamath sees this two-model architecture as the inevitable standard for all AI agents: a real-time voice model for conversation, and an offline LLM for heavy lifting. The startup already counts RingCentral and Truecaller as customers, and Kamath argues that for customer support platforms like Sierra and Decagon, building proprietary voice models is a distraction from their actual business.
Competition is real. ElevenLabs and Cartesia occupy adjacent territory, though Smallest.ai is drawing a hard line around real-time enterprise voice agents — no dubbing, no podcasting, no side quests. Kamath’s North Star is unambiguous: “We want our models to break the Turing test. You should speak to our model and not know it’s AI or human.” Whether a small model can out-charm a well-prompted large one in the wild remains an open question, but $21 million says a lot of people are willing to find out.
