Thinking Machines Lab Announces Interaction Models Preview

Thinking Machines Lab, the artificial intelligence startup founded last year by former OpenAI chief technology officer Mira Murati and researcher John Schulman, announced a research preview of interaction models on Monday. These native multimodal systems enable full-duplex processing, allowing AI to handle audio, video and text inputs and outputs simultaneously for more natural, real-time conversations.

Current AI models operate in a turn-based manner, pausing to process complete user inputs before responding. This creates delays that force users to adapt their communication, such as phrasing queries like emails or batching thoughts. Thinking Machines addresses this with a multi-stream, micro-turn architecture that processes 200-millisecond chunks of data concurrently. The system uses encoder-free early fusion to embed raw audio and image signals directly into a transformer, eliminating reliance on separate encoders.

At the core is TMLInteractionSmall, a 276-billion-parameter mixture-of-experts model with 12 billion active parameters. It pairs a fast interaction model for dialogue management, presence awareness and immediate responses with an asynchronous background model for...