A benchmark suite (one of the few) for measuring interactivity, used to evaluate TML-Interaction-Small. The model is given prerecorded audio and must respond at the correct times. Versions seen:

  • V1 — simple turn-taking latency (seconds). TML-Interaction-Small: 0.40s (best of the compared models).
  • V1.5 — average interaction quality across scenarios (interruption, backchannel, talking to others, background speech). TML-Interaction-Small: 77.8 vs ~45–54 for GPT- realtime and Gemini baselines.
  • V3 — response quality / pass@1 with audio + tools. TML: 82.8 / 68.0 (background agent enabled).
  • Audio MultiChallenge — intelligence / instruction-following (the “intelligence” axis paired against FD-bench’s interactivity axis).
  • QIVD (video+audio QA, streaming), BigBench Audio, IFEval (VoiceBench / text), Harmbench (refusal).
  • Internal proactivity benchmarks: TimeSpeak, CueSpeak, and adapted RepCount-A, ProactiveVideoQA, Charades — where no existing model performs meaningfully.

Example clips (from the source)

The “New dimensions of interactivity” section ships 5 audio-comparison examples, each pitting the input against our model and the baselines: https://thinkingmachines.ai/audio/interaction-models/example-{1..5}/{input,our_model,gpt_realtime_2,gpt_realtime_2_thinking_xhigh,gemini_think_high,gemini_think_minimal}.wav. Example-5 also has a benchmark video:

Full catalog on Interaction Models (TML blog).

Caveats

Some baseline numbers are self-reported via Scale AI / Artificial Analysis, not run by the authors — worth keeping in mind when comparing.

The baseline set is GPT-realtime and Gemini-live; notably absent is Moshi, an open full-duplex model that would be a natural point of comparison (open thread — see Moshi (Kyutai paper)).

Source: Interaction Models (TML blog).