Summary: May 11 saw the release of “Interaction Models,” a lengthy research post from Thinking Machines Lab, Mira Murati’s company, started after her time as OpenAI’s CTO. This post says the weights should handle it. The proposal — a route called interaction models — teaches the model real-time interaction directly in training instead of bolting it together from outside engineering such as voice-activity detection.
May 11 saw the release of “Interaction Models,” a lengthy research post from Thinking Machines Lab, Mira Murati’s company, started after her time as OpenAI’s CTO. The proposal — a route called interaction models — teaches the model real-time interaction directly in training instead of bolting it together from outside engineering such as voice-activity detection. TML-Interaction-Small is the opening research preview — a mixture-of-experts model with 276B parameters in total and 12B active — and it senses and produces video, speech, and text all at once.
Every live voice product on the market bets on the same architecture: full-duplex interaction is handled by the pipeline. This post says the weights should handle it. Teams shipping voice agents, or any product that must respond live, now have a clear signpost for where the stack goes next.
The Case for Training Interactivity into the Model
Every conversational system in service today works in turns: perception shuts off while generation is running, and whatever feels real-time is stitched together from outside machinery — VAD, turn-taking policies, interruption rules. The bitter lesson anchors the lab’s argument: hand-assembled engineering does not scale up as the model gets smarter; only interaction learned as part of the model itself keeps the pace. An interaction model perceives while it generates — overlapping speech, interruptions, and silence all remain in context, and no manufactured turn boundary exists.
How the 200ms Micro-Turn Design Works
At the center sits a mechanism called time-aligned micro-turns: in 200ms units the model alternates between taking in input and producing output, so a conversation’s rhythm survives in full. The model architecture follows an encoder-free early-fusion path: speech becomes dMel tokens through an embedding layer kept lightweight; images get divided into 40x40 patches for an hMLP to encode; audio comes out of a flow head; and every piece is trained jointly from scratch. The systems work is just as serious: streaming inference is built on SGLang (with changes contributed upstream), training and sampling agree bit-for-bit thanks to batch-invariant kernels, and the team hand-built its own communication kernels based on NVLS plus gather+gemv kernels for the MoE.
A Two-Model Split: Live Up Front, Thinking in the Back
The system hands the job to two models. The interaction model stays live at the user’s side; long-horizon agentic tasks, browsing, tool calls, and reasoning are handed to a background model running asynchronously. Both share one context, and its output streams back into the dialogue when the timing fits. Those results show the split works: with the background agent attached, BigBench Audio jumps from 75.7 to 96.5, and FD-bench v3 takes the top score at 82.8/68.0 — reasoning-level intelligence arriving at non-thinking latency.
The Scoreboard: Where It Wins and Where It Trails
Interaction quality on FD-bench v1.5 lands at 77.8 for TML-Interaction-Small, with the field at 46.8 for GPT-realtime-2.0 (minimal), 48.3 for GPT-realtime-1.5, 54.3 for Gemini-3.1-flash-live (minimal), and 39.0 for Qwen 3.5 OMNI. It also records the fastest turn-taking, 0.40 seconds, while the baselines range from 0.57 to 2.14 seconds. The lead is not across the board: 43.4 on Audio MultiChallenge ranks first only in the non-thinking “instant” bracket, and GPT-realtime-2.0 xhigh sits above it at 48.5; on IFEval text its 89.7 trails the 95.2 of GPT xhigh. Refusal on HarmBench reaches 99.0 — a real strong suit. Alongside these, the lab introduced fresh benchmarks on which the baselines barely score: CueSpeak 81.7 vs 2.9, TimeSpeak 64.7 vs 4.3, Charades mIoU 32.4 vs 0, and RepCount-A 35.4 vs 1.3.
Reception, Admitted Limits, and the Road Ahead
Over on Hacker News, the discussion picked up 334 points. The crowd favorite demo showed patience: a person pauses partway through a story for coffee, and the model just sits tight. In the comments, vessenes observed that Opus 4.7 and the GPT-5.x series are each about ten times this model’s size, so scaling room remains; swyx’s caution: the architecture in print is merely the visible tip, with the genuine moat lying in the RL infrastructure and the data recipes behind it; some pushed back that the demo scenarios felt staged and the latency still runs high. The lab is candid about what it cannot yet do: managing context over long sessions, keeping real-time safety alignment on track, coping with weak connectivity, growing the model, and keeping the interaction and background models in step. In the months ahead a limited research preview opens up; bigger models are promised before the year is out; and grants to fund research on interactivity benchmarks are being arranged.
How Developers Should Read This
Three practical consequences. First, voice-product evaluation needs a new yardstick: metrics built for turn-based ASR-TTS stacks cannot capture micro-turn behavior; in time, FD-bench-style testing will be the reference standard. Second, middleware gets squeezed: turn-taking policy and interruption detection form precisely the layer of scaffolding that model providers can fold into the model itself. Third, the 2026 model race is no longer a single-axis intelligence contest — responsiveness in real time has become a competitive dimension of its own, which backs up the start-of-year read on the accelerating model race.