On May 11, Thinking Machines Lab — the research house founded by former OpenAI CTO Mira Murati — published a long piece titled “Interaction Models.” Its proposal: train real-time interactivity into the model itself, instead of assembling it from external scaffolding like voice-activity detectors. The first research preview, TML-Interaction-Small, is a 276B-parameter mixture-of-experts with 12B active parameters, perceiving and emitting speech, video, and text at the same time.
The post takes aim at an assumption shared by every current real-time voice product: full-duplex interaction belongs in the pipeline. Thinking Machines argues it belongs in the weights. For teams building voice agents and live products, that is a pointed signal about where the stack is heading.
Why Interactivity Belongs in the Model
Today’s conversational systems are turn-based: while the model generates, it stops perceiving, and the sense of real time is faked with VAD, interruption rules, and turn-taking heuristics bolted on outside. The lab’s core argument leans on the bitter lesson: this hand-built engineering will not scale with model intelligence. Only interaction trained as a native capability keeps pace. In an interaction model, silence, overlapping speech, and interruptions all stay in context — there are no artificial turn boundaries to trip over.
The 200ms Micro-Turn Architecture
The core mechanism is time-aligned micro-turns: the model processes input and generates output in interleaved 200ms chunks, so the rhythm of a conversation is preserved in context rather than discarded. The architecture uses encoder-free early fusion — audio enters as dMel tokens through a lightweight embedding layer, images are split into 40x40 patches encoded by an hMLP, and a flow head decodes audio output, all co-trained from scratch. The engineering holds up too: streaming inference runs on SGLang (changes upstreamed), batch-invariant kernels keep trainer and sampler bit-for-bit aligned, and the team wrote its own NVLS communication kernels and gather+gemv MoE kernels.
Two Models: Present in Real Time, Thinking in the Background
The system design splits the job. A real-time interaction model stays present with the user, while an asynchronous background model handles reasoning, tool use, browsing, and long-horizon agentic tasks; the two share context, and results stream back into the conversation at the right moments. The numbers say the split works: BigBench Audio rises from 75.7 to 96.5 with the background agent, and FD-bench v3 lands at 82.8/68.0, the best in the comparison — effectively reasoning-model intelligence delivered at non-thinking-model latency.
Benchmarks: Where It Leads and Where It Loses
On FD-bench v1.5 interaction quality, TML-Interaction-Small scores 77.8 against GPT-realtime-2.0 (minimal) at 46.8, GPT-realtime-1.5 at 48.3, Gemini-3.1-flash-live (minimal) at 54.3, and Qwen 3.5 OMNI at 39.0. Turn-taking latency is 0.40 seconds — the best shown, versus 0.57 to 2.14 seconds for the baselines. The lead is not universal: its Audio MultiChallenge score of 43.4 only tops the “instant” (non-thinking) group, with GPT-realtime-2.0 xhigh higher at 48.5; on IFEval text it scores 89.7 versus 95.2 for GPT xhigh. HarmBench refusal, at 99.0, is a genuine strong point. The lab also published a set of new benchmarks where baselines score near zero: TimeSpeak 64.7 vs 4.3, CueSpeak 81.7 vs 2.9, RepCount-A 35.4 vs 1.3, and Charades mIoU 32.4 vs 0.
Reception, Limits, and Timeline
The Hacker News thread collected 334 points. The demo that won people over was patience: a user pauses mid-story to sip coffee and the model simply waits. Commenter vessenes noted the model is roughly one-tenth the size of the Opus 4.7 and GPT-5.x series, leaving room to scale; swyx cautioned that the published architecture is the tip of the iceberg, with data recipes and RL infrastructure as the real moat; critics found the demo scenarios contrived and the latency still a touch high. The lab’s own limitations list is candid: long-session context management, dependence on good connectivity, real-time safety alignment, scaling the model, and tighter interaction-background coordination. A limited research preview opens in the coming months, larger models arrive later this year, and the lab is launching a research grant for interactivity benchmarks.
What It Means for Developers
Three practical effects. First, voice products need new yardsticks: turn-based ASR-TTS pipelines cannot measure micro-turn behavior, and FD-bench-style evaluations are on their way to becoming the reference. Second, middleware value gets squeezed: interruption detection and turn-taking policy layers are exactly the scaffolding a model vendor can absorb. Third, the 2026 model race no longer runs on intelligence alone — real-time responsiveness is now its own competitive axis, which lines up with how the year opened for model competition.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.