Thinking Machines ships first 'interaction model' for parallel voice+video as Meta poaches 7 founders

Mira Murati's Thinking Machines Lab released its first model — a parallel audio/video/text 'interaction model' that processes inputs in 200ms chunks to enable phone-call-style turn-taking, explicitly targeting GPT Realtime 2 and Gemini Live on voice latency and naturalness. The release lands as Meta has poached at least seven founding members, including Mark Jen, Yinghai Lu, and Tinker lead Joshua Gross, who joined Meta Superintelligence Labs.