The interaction model released as a research preview by Thinking Machines Lab. Presented as the first model with both strong intelligence/instruction-following and interactivity.

Overview video — Introducing interaction models:

Capability demos and benchmark clips catalogued on Interaction Models (TML blog).

Architecture

Benchmark snapshot

BenchmarkResultNote
FD-bench V1 turn-taking latency0.40sbest of all models compared
FD-bench V1.5 average77.8next best ~54 (Gemini min)
FD-bench V3 response quality / pass@182.8 / 68.0with background agent enabled
Audio MultiChallenge APR43.4%best instant model; < GPT-2.0 xhigh (48.5%)
IFEval (text)89.7%
Harmbench refusal rate99.0%

Limitations

Larger pretrained models from the same family are currently too slow to serve in this real-time regime; long A/V sessions strain context management; needs reliable connectivity.

Training data

TML does not disclose the training corpus. The blog says only that the model is trained from scratch on continuous audio + video + text, and that capabilities “improve in quality as we scale up model size and training data” (the bitter-lesson bet). The only data specifics are for safety, and they’re synthetic: a TTS model generates colloquial refusal / over-refusal examples, and an automated red-teaming harness generates multi-turn refusal data to keep speech refusals in parity with the text model.

So how real-world data trains the core interaction behaviour is unknown for TML-small. The closest documented analogue is Moshi (see its “Training data” section): unlabeled real audio + ASR-generated aligned text, real two-channel conversations (Fisher) for full-duplex, and diarization to synthesize multi-stream structure — principles that plausibly transfer, with TML additionally needing time-aligned video that Moshi’s audio-only recipe doesn’t cover.

Compared to Moshi

Moshi (Kyutai, 2024) is the closest predecessor — also real-time and full-duplex, but audio-only, a single model (no background split), and built on a discrete codec rather than Encoder-Free Early Fusion. TML-Interaction-Small extends the regime to audio + video + text with stronger intelligence. See Moshi (Kyutai paper).

Detailed in source: Interaction Models (TML blog).