Thinking Machines Lab · 2026-05-11 · major
Thinking Machines TML-Interaction-Small — 276B-A12B Native Interaction Model Listens, Talks, and Watches Concurrently at 0.40s Turn-Taking Latency
Mira Murati's lab unveils a 276B mixture-of-experts model with 12B active parameters that fuses audio, video, and text in 200ms microturns, paired with an asynchronous background model for slow reasoning. Research preview opens later this year.
A 276B mixture-of-experts that holds a real-time conversation while a separate background model does the slow reasoning.
Key specs
| Parameters | 276B |
|---|---|
| Active params | 12B |
| Turn taking latency | 0.40s |
| Fd bench v1.5 quality | 77.8 |
| Time speak accuracy | 64.7% |
| Cue speak accuracy | 81.7% |
What is it?
Thinking Machines Lab's first public model family. TML-Interaction-Small is a 276B MoE (12B active) trained to handle voice, video, and text as one continuous stream rather than a turn-based chat. A second 'background model' handles tool calls, web search, and complex reasoning, then streams results into the live conversation as they arrive.
How does it work?
The model uses encoder-free early fusion of audio and image tokens in time-aligned 200ms microturns, similar in spirit to Meta's Chameleon. It can speak, listen, watch, and react concurrently — including barge-in, silence handling, and timed cues — with no separate voice-activity detector. Two new in-house benchmarks (TimeSpeak, CueSpeak) measure when a model should speak, not just what it should say.
Why does it matter?
Today's voice assistants stitch ASR, an LLM, and TTS together and stall while one of them thinks. A native interaction model collapses that stack and pushes turn-taking latency to about 0.40s, close to a human phone call. It is also Thinking Machines' first concrete product signal — and a direct shot at OpenAI's Realtime and Google's Gemini Flash Live.
Who is it for?
voice and agent product teams
Try it
Limited research preview via research-preview@thinkingmachines.ai