Thinking Machines Lab, the AI startup founded last year by former OpenAI CTO Mira Murati, just announced what it calls interaction models. According to TechCrunch AI, the company unveiled the approach on Monday with a research model named TML-Interaction-Small. The pitch is simple: an AI that can listen and speak at the same time, the way humans actually talk.
What’s new
Every chatbot you’ve ever used runs on turn-taking. You send, it replies. You wait, it waits. Thinking Machines wants to kill that rhythm. Its model processes your input and generates a response simultaneously, a setup engineers call “full duplex.” Think phone call, not text thread.
The headline number: TML-Interaction-Small responds in 0.40 seconds. That’s roughly the cadence of natural human conversation, and TechCrunch AI reports it beats comparable systems from OpenAI and Google on latency.
Why this matters
Voice AI has been chasing this problem for years. OpenAI’s Advanced Voice Mode, Google’s Gemini Live, and a handful of startups all tried to make conversations feel less robotic. Most still stumble on the same thing: they can’t gracefully handle interruptions, overlapping speech, or the small “mhm” sounds that signal you’re still listening.
Building interactivity into the model itself, rather than gluing it on top of a text-based system, is a different architectural bet. If it works, the downstream effects could be significant:
- Voice agents that handle real customer calls without awkward pauses
- Tutoring and coaching apps where the AI can jump in to correct mistakes
- Accessibility tools that keep pace with natural speech
- Live translation that doesn’t lag a full sentence behind
The catch
This is a research preview, not a product. Nobody outside the company can touch it yet. Thinking Machines says a limited preview lands in the coming months, with broader access later this year.
Benchmarks on paper are one thing. Real conversations are messy. Background noise, accents, people talking over each other, the AI needing to know when to shut up. None of that shows up in a latency number.
The bigger picture
Murati’s startup has been quiet since launch, raising eyebrows with a reported multi-billion-dollar valuation and very little public output. This is the first concrete technical claim from the team, and it’s a swing at one of the harder problems in voice AI.
If TML-Interaction-Small holds up outside the lab, expect competitors to follow fast. Full-duplex voice is the kind of feature that becomes table stakes the moment one company nails it. Watch for the preview drop in the coming months. More details at the original source.