Thinking Machines Lab releases its first AI model with native interactivity, claiming it outperforms OpenAI's GPT-Realtime-2 and Google's Gemini Live on interaction quality

Thinking Machines Lab, founded in February 2025 by Mira Murati and other former OpenAI researchers, published a research preview of TML-Interaction-Small, a 276-billion-parameter mixture-of-experts model with 12 billion active parameters designed to handle voice, video, and text interaction natively.

The model processes audio and video directly in parallel 200-millisecond chunks rather than relying on separate components like voice activity detectors. This time-aligned approach lets the model decide whether to stay silent, interject, or speak alongside the user, paired with a background reasoning model that handles longer tasks while the interaction model keeps the conversation flowing.

On FD-bench v1.5 (which measures interaction quality), TML-Interaction-Small significantly outperforms both GPT-Realtime-2 and Gemini-3.1-flash-live. Response latency is 0.40 seconds, compared to 1.18 seconds for GPT-Realtime-2 (minimum) and 0.57 seconds for Gemini. On the lab's own benchmarks for time awareness and visual proactivity, no existing model can meaningfully perform these tasks.