TML-Interaction-Small

TML-Interaction-Small is an interaction model unveiled on 11 May 2026 by Thinking Machines Lab as a research preview. It uses a Mixture-of-Experts architecture with 276B total parameters and 12B active parameters. The model processes continuous audio, video and text streams in 200 ms micro-turns, generating text and audio concurrently without external voice-activity-detection components.

The architecture uses encoder-free early fusion: audio is represented as dMel, images are split into 40×40 patches encoded by an hMLP, and the audio decoder uses a flow head. All components are co-trained from scratch with the transformer. On FD-bench V1 the model reaches 0.40 s turn-taking latency, and 43.4% on Audio MultiChallenge APR. The system pairs with an asynchronous background model handling longer reasoning and tool use.