Thinking Machines has released a research preview of interaction models—AI systems that handle real-time collaboration natively rather than through external scaffolding. Unlike traditional turn-based interfaces where models wait for users to finish typing before responding, these new models maintain continuous two-way exchange across audio, video, and text. The company argues interactivity has been treated as an afterthought in AI development.

The research addresses what Thinking Machines calls a "bandwidth bottleneck" in human-AI collaboration. Current frontier models push humans out of the loop not because the work doesn't need them, but because interfaces have no room for them. One cited model card acknowledged that users perceived models as "too slow" in interactive "hands-on-keyboard" patterns, prompting concerns about collaboration limitations.

The interaction model uses a time-aligned micro-turn architecture, processing continuous input and output streams rather than alternating token sequences. This enables capabilities like implicit dialog management, verbal and visual interjections, simultaneous speech, and concurrent tool calls while conversing. An asynchronous background model handles deeper reasoning tasks while the interaction model maintains real-time presence.

The system shares context between the real-time interaction model and the background reasoning model, creating an experience that feels more like collaborating with a person than prompting a machine. This approach represents a fundamental shift in how AI systems are designed to engage with users.