Thinking Machines Unveils Interaction Models

Thinking Machines has released a research preview of Interaction Models.

• The model aims for interaction that simultaneously takes in audio, video, and text, responds in real time, and intervenes immediately when needed.
• The core design is a time-aligned approach that stitches together input and output in 200ms micro-turn units.
• Instead of a separate harness, the model handles conversational flow, interruptions, simultaneous speech, and visual signals internally.
• A background model handles long-term reasoning and tool use, while the interaction model continuously reflects those results back into the conversation.
• The released TML-Interaction-Small is stated to be a 276B MoE with 12B active parameters.
• Interactivity and intelligence were evaluated together via FD-bench v1.5 and Audio MultiChallenge, presenting a turn-taking latency of 0.40 seconds.
• The company previewed a limited research preview opening in the coming months, and announced larger models and broader availability within the year.