Thinking Machines details its "interaction models"

Thinking Machines Lab unveils TML-Interaction-Small, an interaction model outperforming GPT-realtime-2.0 and Gemini 3.1 Flash Live on the FD-bench v1.5.

The lab founded by Mira Murati is publishing a research paper on interaction models, models that natively integrate interactivity instead of simulating it via an external harness. The principle: continuously process audio, video, and text streams in 200-millisecond micro-turns, which allows the system to listen and speak simultaneously, react to visual cues without explicit prompting, or intervene at the right moment in the flow of an exchange.

The architecture combines two components. A real-time interaction model manages the conversation and perceives in parallel with its response. A background, asynchronous model handles long-form reasoning, tool calls, and web navigation, then re-injects its results into the dialogue. Audio and video undergo early fusion without a dedicated encoder (dMel for sound, 40x40 patches for image), with end-to-end co-training.

The released version, TML-Interaction-Small, is a mixture-of-experts model with 276 billion parameters, 12 billion of which are active. On FD-bench v1.5, which evaluates interaction quality (interruption management, backchannels, background speech), it outperforms GPT-realtime-2.0 and Gemini 3.1 Flash Live, while remaining competitive on Audio MultiChallenge, a pure intelligence benchmark. For Thinking Machines Lab, this audio-video-text continuity is the condition for transitioning from turn-by-turn exchange to true human-AI collaboration.

A limited research preview will open in the coming months, with a broader release planned for the end of the year. A fellowship program is open in parallel to support research on interactivity evaluation.