Thinking Machines announced a new SOTA realtime voice model

Thinking Machines unveiled a research preview of new interaction models for continuous, native multimodal conversation.

The system uses a micro-turn design with a time-aware interaction model plus a background reasoning model to manage speech, timing, and tools in parallel.

Early results reportedly outperform existing models on both intelligence and interaction quality.

Access is currently limited to selected researchers, with broader availability planned later this year.