Founded by Albert Gu and Karan Goel out of the Stanford lab where Mamba was written. The bet is that state-space models, which keep a fixed-size state instead of attending over every past token, are the right substrate for anything real-time or on-device. The product is voice (Sonic), but the research line is the architecture. A $100M round was widely reported in October 2025 alongside Sonic-3; the company has not published an announcement, so it is not counted here.