Thinking Machines Lab unveils interaction models to enable interruptible, full‑duplex AI
Thinking Machines Lab unveils “interaction models”—a full‑duplex AI that can interrupt users, claiming human‑speed 0.40s responses in a limited research preview.
Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, announced a new class of models it calls interaction models that are designed to process and generate speech simultaneously. The company says these models operate in a full‑duplex mode, letting the AI interject or respond without waiting for a complete user turn. The announcement positions interaction models as a shift toward more conversational, phone‑call‑style exchanges rather than the ask‑then‑reply format familiar to users today.
Thinking Machines announces interaction models
Thinking Machines Lab framed the new capability as a move to make AI more interactive and natural in conversation. The company highlighted a model named TML‑Interaction‑Small as the initial proof of concept for the approach. Mira Murati, who founded the startup last year, has led the effort to embed interactivity at the model level rather than layering it on top of existing chat architectures.
The lab described interaction models as research rather than finished products and emphasized that the announcement is an early step. Company materials accompanying the reveal include performance benchmarks and technical explanations of the full‑duplex architecture. Thinking Machines is positioning the work as a platform innovation that could be applied across voice assistants, live support, and collaborative AI tools.
Full‑duplex design aims to make AI interruptible
Full‑duplex refers to the ability to send and receive information at the same time, a capability common in human phone conversations but rare in deployed AI systems. Traditional models operate in half‑duplex: a user speaks or types, then the AI responds after processing the full input. Thinking Machines’ interaction models are designed to begin generating output while still ingesting user input, enabling mid‑turn responses and interruptions.
The company argues this design will make interactions feel faster and more natural, reducing latency and mirroring human conversational rhythms. Implementing true full‑duplex processing requires novel engineering to manage overlapping input and output streams and to reconcile partial user utterances with model predictions. Those technical hurdles are central to how the startup describes its research.
Benchmark claims and latency comparisons
Thinking Machines reported that TML‑Interaction‑Small can respond in roughly 0.40 seconds, which the company describes as comparable to natural human conversational speed. The lab contrasted that figure with response times it said were typical of comparable systems from major AI vendors, implying a material speed advantage. The benchmarks were presented in the company’s research materials alongside explanations of test conditions and metrics.
Observers caution that benchmark claims are useful but not definitive until systems are tested in diverse real‑world settings. Latency measured in controlled tests can change when models face noisy audio, simultaneous speakers, or complex conversational turns. Thinking Machines’ numbers are notable, but independent evaluations will be needed to confirm performance across practical deployments.
Limited research preview and release timing
Thinking Machines said it plans a limited research preview in the coming months, followed by a broader release later in the year. The company characterized the initial availability as restricted to researchers and partners to gather feedback and iterate. There are no immediate plans to ship the technology as a consumer product in its current form, according to the announcement.
That staged rollout is consistent with many AI labs’ approach to complex, potentially disruptive features: release a controlled preview, refine the model based on usage and safety assessments, then widen access. The lab has not published a firm public schedule with exact dates or partner names for the preview, leaving timing and scope open until nearer to launch.
Potential applications in live voice and augmented workflows
Interaction models could change how voice assistants, customer service bots, and collaborative tools behave by enabling interruptible responses and more dynamic turns. In customer support, for example, an agent‑assist model could interject suggestions while a human is speaking, or an assistant could correct misunderstandings in real time. Developers building multi‑party voice systems may also use the approach to better manage interruptions and turn taking.
The technology may enable new product designs that blend AI and human contributions more tightly during meetings, calls, and live broadcasts. However, product teams will need to redesign interfaces to give users clear control over when and how the AI interrupts, and to manage expectations about timing and content of interjections.
Safety, privacy and open questions
Making AI interruptible raises safety and privacy questions that Thinking Machines acknowledges as subjects for further study. Interruptions could be intrusive if not properly signaled, and overlapping audio streams complicate transcription and logging policies. There are also risks around misaligned interjections—instances where a model guesses intent prematurely and produces misleading or incorrect information.
Experts will look for details on consent mechanisms, opt‑out controls, and transparency features in the research preview. The lab has signaled an intention to gather feedback from researchers and partners before wider distribution, but many questions remain about moderation, user controls, and handling of ambiguous speech.
Thinking Machines Lab’s interaction models represent a technical pivot toward lower latency, conversationally fluent AI that can intervene mid‑turn. The company’s 0.40‑second latency claim and the full‑duplex architecture are notable advances in research, but practical impact will depend on how well prototypes perform outside controlled tests and how product teams address safety and usability trade‑offs. The upcoming limited research preview will provide the first tangible tests of whether interruptible, phone‑call‑style AI can deliver meaningful improvements in real‑world conversations.
