Interaction Model
#13 in Reasoning Modelsthinking-machines-lab · since 11. Mai 2026 (Research Preview) · 13× · last seen Jun 30, 2026
28
Momentum
TML-Interaction-Small is the first in-house AI model from Thinking Machines Lab (founded by former OpenAI CTO Mira Murati), announced on May 11, 2026 as a research preview. It is a 276-billion-parameter mixture-of-experts model (12B active parameters) that natively and simultaneously processes audio, video, and text in 200ms micro-turns — without external voice-activity-detection scaffolding. The model operates in full-duplex mode, able to listen and speak at the same time, and react proactively to visual cues. A separate asynchronous background model handles deeper reasoning and tool-use tasks.
Momentum trend
04.04.03.07.
Features
| Key Benchmark (%) | FD-bench V1.5 (Interaction Quality): 77.8 vs. 46.8 (GPT-Realtime-2.0) | FD-bench V1 (Latency): 0.40 s vs. 1.18 s (GPT-Realtime-2.0) | BigBench Audio (with background model): 96.5% | Audio MultiChallenge: 43.4% |
| Context Window (Tokens) | Not officially disclosed; continuous 200 ms micro-turn streams; long sessions reportedly require active context management per vendor |
| License | Not publicly documented (research preview, gated access; no open-source or commercial license communicated) |
| Multimodality | Audio (input/output), video (input), text (input/output); full-duplex; proactive visual response without explicit prompt; live translation; encoder-free early fusion (dMel audio, hMLP video patches) |
| Platform | Closed research preview via thinkingmachines.ai; no public API/SDK at launch; access via application by email (interaction@thinkingmachines.ai) |
| Price per 1M Tokens | Not publicly disclosed (no commercial access at launch) |
| Release Date | May 11, 2026 (research preview; public release announced for later in 2026) |