Anyone who has ever spoken with a voice assistant knows that moment. You finish your sentence, silence follows, there is a brief wait, and then the answer comes. The whole thing feels like an old satellite phone call. Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, has announced that it wants to change this.
The company published a research paper on an approach it calls “interaction models.” These are models that process input and generate a response simultaneously, rather than sequentially. Technically, this principle is called full duplex, and in plain language it means that artificial intelligence does not have to wait for you to finish speaking before it can begin responding.
Today’s models work differently. The user speaks and the model listens. The model begins responding, and the user remains silent. This mode of communication works, but it is artificial. No real conversation works this way. People interrupt each other, correct one another mid-sentence, and react to the other person’s facial expression before a word is even spoken. Until now, AI has not been able to do this.
Interaction Model
The model does not work with entire sentences or paragraphs as input and output, but with what are known as 200-millisecond micro-turns. Every 200 ms, it processes a piece of input while simultaneously generating a piece of output. To the user, it looks like a fluid conversation; for the model, it is a constant alternation between listening and speaking in tiny segments.
To make the entire system work quickly enough, the team also redesigned how the model receives audio and video. Instead of the large pretrained encoders used by most multimodal models, they opted for a lightweight processing system applied directly during training. Audio arrives as dMel spectrograms and video as 40×40-pixel patches. Everything is trained from scratch as a single system.
There is one more trick. Because a fast model cannot also be as intelligent as slower systems designed for deep reasoning, Thinking Machines added a second, slower model running in the background. The interaction model handles the conversation, but as soon as it encounters a complex task, it passes it to the background model. That model works asynchronously and feeds the results back into the ongoing conversation. In the meantime, the user does not have to stop talking to the assistant; smarter responses are simply incorporated as they become available.
The result is a model called TML-Interaction-Small. Its latency—the delay between the end of the user’s speech and the start of the model’s response—is 0.40 seconds. For comparison, GPT-realtime-2.0 in its minimal configuration requires 1.18 seconds, while Gemini-3.1-flash-live takes 0.57 seconds. Natural human conversation falls within roughly the same range as TML’s result.
On FD-bench V1.5, a benchmark that measures interaction quality in situations such as user interruptions, simultaneous speech, or responses to visual cues, TML-Interaction-Small achieved a score of 77.8, while GPT-realtime-2.0 scored only 46.8 and Gemini-3.1-flash-live 45.5. That is a significant difference. The model also has a total of 276 billion parameters, of which it actively uses 12 billion. It employs an MoE (mixture of experts) architecture, which activates only a subset of parameters for each input.
Demonstrations show capabilities that existing voice assistants do not have at all. The model continuously monitors video and, without being prompted, warns a person at a computer when they sit with their back hunched. It can translate live while speaking at the same time as the user, rather than after them. It counts push-ups during exercise directly from video. It launches a web search in the middle of a conversation and naturally weaves the results into its response.
This is enabled by its ability to respond to visual cues without a verbal prompt. Existing systems, including GPT-realtime-2.0, wait for an audio signal. If no signal arrives, they remain silent. On the internal Charades benchmark, where the model must estimate the beginning and end of an action in a video based solely on visual signals, TML-Interaction-Small achieved an mIoU score of 29.4. GPT-realtime-2.0 scored practically zero.
It Is Not as New as It Seems
Researcher Sean Goedecke notes in his analysis that full-duplex voice models have existed for some time. Moshi from the French laboratory Kyutai, PersonaPlex from Nvidia, and Nemotron-VoiceChat operate on the same micro-turn principle. Thinking Machines did not invent anything in this regard.
What is different, however, is the scale. TML-Interaction-Small has approximately forty times as many total parameters as Moshi and roughly twice as many active parameters. And it adds video. None of the existing full-duplex models support video, probably because they simply were not powerful enough. According to Goedecke, this is the real technical contribution: a full-duplex model at scale and with visual input.
Adding a background “smarter” model improves benchmark results, but it also raises questions. Will the interaction model be smart enough to know when to hand off a task? Will responses feel disjointed when background results arrive with a delay? That will only become clear in real-world use.
Availability
For now, no one can try it. Thinking Machines has announced limited research access within the next few months and plans a broader launch by the end of this year. This is therefore not a product, but a research demonstration with specific figures and demos.
The company was founded in 2025 after Mira Murati left OpenAI. Since then, it has attracted a number of researchers from OpenAI, Meta, and Anthropic and signed computing infrastructure agreements with Google Cloud and Nvidia. According to available reports, its current valuation is around 12 billion dollars.
Sources: techcrunch.com and jawlah.co
