Thinking Machines Interaction Models May 11, 2026
- FD-bench v1 turn-taking latency, seconds, streaming audio 0.4
- FD-bench v1.5 average quality, streaming audio 77.8
- FD-bench v3 Pass@1, with background agent, audio plus tools 68
- QIVD accuracy, streaming video plus audio 54
- Audio MultiChallenge APR, audio 43.4
- BigBench Audio accuracy, audio, first reported value 75.7, alternative 96.5 setting shown 75.7
- IFEval (VoiceBench) accuracy, audio 82.1
- IFEval accuracy, text 89.7
- HarmBench refusal rate, text 99
- TimeSpeak macro-accuracy, internal 64.7
- CueSpeak macro-accuracy, internal 81.7
- RepCount-A off-by-one score, internal 35.4
- ProactiveVideoQA PAUC at omega 0.5, internal 33.5
- Charades mIoU, internal visual proactivity adaptation 32.4