TML Interaction
#32 en Modèles de reasoningThinking Machines · depuis 11. Mai 2026 (Forschungsvorschau/Announcement) · 19× · vu le 30 juin 2026
TML-Interaction-Small is the first publicly announced model from Thinking Machines Lab (Mira Murati's startup), unveiled as a research preview on May 11, 2026. It is a 276-billion-parameter Mixture-of-Experts model with 12 billion active parameters that natively processes audio, video, and text in continuous 200-millisecond "micro-turns" rather than waiting for traditional conversational turns. Unlike prior voice AI systems, it requires no external voice-activity-detection harness, instead building interactivity directly into the model, enabling simultaneous listening, seeing, and speaking as well as natural interruptions. The model is currently available only to select partners in a limited research preview, with a wider rollout planned for later in 2026; pricing has not yet been announce
Fonctionnalités
| Key Benchmark (%) | FD-bench v1.5 (interaction quality): 77.8 vs. 54.3 (Gemini-3.1-flash-live minimal) and 46.8 (GPT-realtime-2.0 minimal) |
| License | No open access/no open weights; gated Research Preview, access only upon request via interaction@thinkingmachines.ai |
| Multimodality | Input: continuous audio, video/images, text; Output: audio and text – all modalities processed simultaneously in 200ms micro-turns, encoder-free early-fusion architecture |
| Platform | Cloud-based Research Preview (select partners only); serving optimizations upstreamed to, among others, the SGLang inference framework |
| Price per 1M Tokens | Not yet published; pricing to be announced only at wider release |
| Release Date | May 11, 2026 (blog post 'Interaction Models: A Scalable Approach to Human-AI Collaboration'); limited Research Preview to follow in coming months, wider release later in 2026 |