Thinking Machines Lab Cracks Sub-Second AI Conversation With 0.40s Turn Latency

Mira Murati's Thinking Machines Lab announced TML-Interaction-Small, a 276B-parameter MoE research preview that achieves 0.40s turn latency in voice conversation via a full-duplex, multi-stream, micro-turn architecture. The model treats interaction as a first-class citizen of the architecture, not an external software harness. Here is the design, the benchmarks, and the implications for the voice AI race.

Hey guys, Mr. Technology here.

Thinking Machines Lab — the Mira Murati and John Schulman startup that raised $2B at a $12B valuation — published a research preview on May 13, 2026 that re-opens a question the AI industry had quietly given up on. Can a foundation model have a real-time voice conversation with a human at near-human turn latency? The answer from Thinking Machines: yes, at 0.40 seconds per turn, with a 276B-parameter mixture-of-experts model that treats interaction as a first-class citizen of the model architecture, not a software harness bolted on top.

The model is called TML-Interaction-Small. It is not yet publicly available. The company's announcement blog post and the VentureBeat coverage lay out the design and the latency numbers. The architecture is the part that matters, because it is a different answer to a question every other AI lab has been solving with the same scaffolding.

The Problem: Turn-Based Models Are Not Conversational

The current generation of AI voice assistants — including the best from OpenAI, Google, and others — all share the same architectural choice. The model is a text-in, text-out transformer. The voice stack is a software pipeline around the model: speech-to-text model → text model → text-to-speech model. Each stage has latency. The cumulative latency is the user's perceived turn time.

The result: a typical voice assistant has a 1.0-1.5 second turn latency. The model hears a complete utterance, the STT model transcribes, the text model generates a response, the TTS model speaks the response. The user waits. The conversation feels like two people shouting into a canyon.

This is the limit of the turn-based model. The model waits for a complete input, processes, generates, and the user waits for a complete output. There is no overlap. There is no backchannel. There is no "mm-hmm" while the model is thinking. The human in the conversation has been forced to contort themselves to the model's interface.

A real human conversation has none of those properties. The latency is ~200ms. The participants overlap, backchannel, and interject. The conversation is full-duplex. The model architecture, as it currently exists, is the wrong tool for the job.

The Fix: Multi-Stream, Micro-Turn, Full-Duplex

Thinking Machines' answer is a model architecture that processes input and output simultaneously, in 200ms chunks, with no freezing between turns. The key design choices:

Multi-stream, micro-turn design. Instead of waiting for a complete input, the model processes 200ms chunks of audio and video as they arrive, and generates 200ms chunks of output at the same time. The model is listening, thinking, and speaking in parallel. The latency floor is the chunk size, not the cumulative pipeline.

Encoder-free early fusion. Conventional voice stacks use a separate Whisper-class speech encoder. TML fuses audio directly into the transformer as dMel spectrograms, and image patches as 40x40 pixel tokens. The model co-trains all modalities from scratch, with no separate encoder, no separate TTS head, and no software pipeline between the modalities.

Dual model system. The interaction model itself is one half. The other half is a background model — an asynchronous agent that handles long-running reasoning, web browsing, complex tool calls — which streams results back to the interaction model. The interaction model weaves the background results into the conversation as they arrive. The user perceives a model that can do long-running work without breaking the conversational flow.

TML-Interaction-Small specs. 276B total parameters, 12B active per token, MoE with 128 routed experts. Trained from scratch on the multi-stream, micro-turn design. The "small" in the name is aspirational — this is a research preview of a 276B model, with larger and smaller variants planned.

The architecture, in the company's framing, makes interaction a "first-class citizen of model architecture rather than an external software harness." That sentence is the entire pitch. The voice stack is not a software pipeline around a text model. The model is, from the ground up, an interaction model.

The Numbers

Per the VentureBeat coverage and the Thinking Machines blog post:

0.40 second turn latency in voice conversation. Down from the typical 1.0-1.5 seconds of stacked voice assistants.
Sustained overlap and backchannel. The model can interject, say "mm-hmm," ask a clarifying question, and respond to a visual cue (e.g., a user writing a bug in a code snippet on screen) without breaking the conversation.
State-of-the-art combined intelligence and responsiveness on third-party benchmarks. The model does not sacrifice reasoning quality for latency. It is competitive with frontier text models on standard benchmarks while delivering the latency numbers.
The research preview is not yet public. The model is being shown in a research preview video. Limited research preview will open "in the coming months" with wider release "later this year."

The honest read: this is a research preview, not a product. The latency numbers are real. The architectural choices are real. The model is not yet available for testing. The next 6-12 months will determine whether the architecture scales to production and whether the latency numbers hold up under real-world conditions.

Why It Matters: The Voice AI Race Just Changed

The voice AI market in 2026 is dominated by two architectures:

The turn-based stack. OpenAI's GPT-Realtime, Google's Gemini Live, the open-source Moshi, PersonaPlex, and Nemotron VoiceChat. All use the text-model-plus-pipeline architecture. Latency floor: 1.0-1.5 seconds.
The full-duplex stack. TML-Interaction-Small, and the smaller-scale research models that have been published in the last 18 months. Latency floor: 0.4-0.6 seconds.

The full-duplex stack is the right architecture for human conversation. The turn-based stack is the wrong architecture that won because it was the only one that worked with the existing text models. Thinking Machines is now the first well-funded lab to ship a frontier-scale full-duplex model.

The implications:

For OpenAI and Google: the voice stack is now a generation behind. GPT-Realtime and Gemini Live are excellent products, but the architectural floor is higher than what the current pipeline can deliver. The next model release from each lab will need to either adopt the full-duplex architecture or build a hybrid that approximates it.

For the open-source community: the architecture is publishable. Moshi, PersonaPlex, and Nemotron VoiceChat are all smaller-scale open-source voice models. The full-duplex design is implementable. Expect a wave of open-source full-duplex models in the next 6-12 months as the community digests the Thinking Machines paper.

For voice AI product builders: the latency floor just dropped by 60-70%. Products that were marginal at 1.0-1.5 second latency — real-time translation, accessibility tools, in-call coaching, conversational search — become viable at 0.4 second latency. The product surface expands.

For the broader AI race: this is the first major architectural move that did not come out of OpenAI, Anthropic, or Google. Thinking Machines is now a serious competitor for the next product cycle. The $12B valuation looks, in retrospect, like the discount.

The Take

Three things to act on this week.

If you build voice AI products: the latency floor is about to drop. The vendors you are using will either ship a full-duplex option in the next 12 months or get displaced by vendors who do. Start evaluating full-duplex models now, even if they are research previews. The product surface that becomes viable at 0.4s latency is real, and the products that own that surface will be the ones that ship in 2027.

If you invest in AI infrastructure: the voice AI race is a generation-level architectural shift, and the right architecture is now known. The labs that ship full-duplex models will own the next product cycle. The labs that don't will be racing a commoditizing pipeline.

If you are a Murati/Schulman skeptic: the skepticism was reasonable when the company was announced in February 2025. The research preview is the answer. The architecture is the right one, the latency numbers are real, and the team that built this is the team that built the foundation models that defined the last cycle. The bet is paying off.

The 0.40 second number is the headline. The architectural commitment — interaction as a first-class citizen of the model — is the news. The next product cycle for voice AI will be built on this architecture.

— Mr. Technology

Sources: Thinking Machines Lab — Interaction Models: A Scalable Approach to Human-AI Collaboration, VentureBeat — Thinking Machines shows off preview of near-realtime AI voice and video conversation, VentureBeat — Ex-OpenAI CTO Mira Murati unveils Thinking Machines, METR — Measuring AI Ability to Complete Long Tasks, Anthropic Claude Opus 4 model card reference.