<p>Thinking Machines Lab, the company founded by former OpenAI CTO Mira Murati, has announced what it calls “interaction models” — a new family of AI systems designed to handle real-time, multimodal conversation natively rather than bolting interactivity on top of a standard language model.</p>
<p>The core idea is full-duplex communication: the model processes audio, video, and text input while simultaneously generating a response. Most current voice AI works in turns — you speak, it listens, it responds. This approach runs everything concurrently, which means the model can handle interruptions, respond to visual cues mid-sentence, and use tools in the background without pausing the conversation.</p>
<h2>How it works</h2>
<p>The architecture uses 200ms micro-turns, encoder-free early fusion, and a dual-model design. A real-time interaction model handles the live conversation while a separate asynchronous reasoning model works in the background on more complex tasks. The two run in parallel rather than sequentially.</p>
<p>Their first model, TML-Interaction-Small, is a 276B mixture-of-experts model with 12B active parameters. It reportedly achieves a 0.40-second response latency — roughly in line with natural human conversation pace, and faster than comparable models from OpenAI and Google according to the company’s benchmarks on FD-bench.</p>
<p>Thinking Machines also introduced three internal benchmarks — TimeSpeak, CueSpeak, and a visual proactivity task — to measure capabilities they say no existing commercial model can currently perform, including responding to visual interjections and handling simultaneous speech.</p>
<h2>What’s different here</h2>
<p>The distinction the company is drawing is between interactivity as a feature versus interactivity as a foundation. Current real-time voice products from major labs are largely standard language models with audio pipelines wrapped around them. Thinking Machines is arguing that approach has a ceiling — that genuine responsiveness requires the model to be built around interaction from the start.</p>
<p>Whether that architectural bet pays off is still an open question. The research preview is limited, and the wider release is expected later this year. But the benchmark numbers are at least interesting enough to watch.</p>


Collections

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, has announced 'interaction models' — AI systems built natively for real-time, multimodal conversation using full-duplex communication. Unlike current voice AI that operates in turn-based fashion, these models process audio, video, and text simultaneously while generating responses. The architecture uses 200ms micro-turns, encoder-free early fusion, and a dual-model design pairing a real-time interaction model with a background async reasoning model. Their first model, TML-Interaction-Small, is a 276B mixture-of-experts model with 12B active parameters achieving 0.40-second response latency. The company claims this outperforms comparable models from OpenAI and Google on their internal FD-bench benchmark. A limited research preview is available, with wider release expected later this year.

Thinking Machines Lab previews 'interaction models' for real-time multimodal AI