Thinking Machines shows off preview of near-realtime AI voice and video conversation with new 'interaction models'

2026-05-11 · Source: VentureBeat · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Advanced, long

Summary

Thinking Machines, an AI startup founded by former OpenAI CTO Mira Murati and researcher John Schulman, has announced a research preview of "interaction models," a new class of native multimodal systems. These models prioritize interactivity within their architecture, moving beyond the traditional "turn-based" AI interaction. The core innovation is a "full-duplex" architecture that processes 200ms chunks of input and output simultaneously across text, imagery, audio, and video. This enables real-time responses, backchanneling, and proactive interjections. The preview introduces TML-Interaction-Small, a 276-billion parameter Mixture-of-Experts (MoE) model, which features a dual system with an Interaction Model for immediate exchanges and a Background Model for asynchronous reasoning. Benchmarks like FD-bench show TML-Interaction-Small achieving a turn-taking latency of 0.40 seconds and an interaction quality score of 77.8, significantly outperforming competitors like Gemini-3.1-flash-live and GPT-realtime-2.0.

Key takeaway

For CTOs and VPs of Engineering evaluating next-generation AI integration, Thinking Machines' interaction models represent a significant leap beyond current turn-based systems. Their ability to handle simultaneous multimodal input and output, demonstrated by a 0.40-second turn-taking latency, could fundamentally transform enterprise applications requiring natural, real-time human-AI collaboration. You should monitor their upcoming limited research preview for potential pilot programs, especially for use cases in customer service, industrial monitoring, or any scenario where sub-second responsiveness and proactive AI engagement are critical.

Key insights

Thinking Machines' interaction models enable real-time, full-duplex human-AI communication by processing multimodal inputs and outputs simultaneously.

Principles

• Interactivity as a first-class citizen
• Multi-stream, micro-turn processing
• Dual model system for real-time and background tasks

Method

The system uses a multi-stream, micro-turn design, processing 200ms chunks of input/output simultaneously. It employs encoder-free early fusion, taking raw audio (dMel) and image patches (40x40) through a lightweight embedding layer, co-training all components within a transformer.

In practice

• Real-time safety monitoring in manufacturing
• Proactive customer service with natural latency
• Time-aware agents for industrial maintenance

Topics

• Thinking Machines
• Interaction Models
• Full-Duplex AI
• Multimodal AI
• Low-Latency AI

Best for: Machine Learning Engineer, CTO, VP of Engineering/Data, AI Scientist, AI Engineer, Director of AI/ML

Related on AIssential

• Thinking Machines drops a new, highly responsive model designed for humanlike interactions in real time — New AI architecture enables full-duplex, real-time multimodal interaction by processing data in 200-millisecond micro-turns.
• New Trends in AI Interaction Models — Thinking Machines Lab's Interaction Models enable full-duplex AI communication, allowing simultaneous listening and talking to eliminate conversational latency.
• New AI Model SHOCKED Me — Thinking Machines Labs' new model offers real-time, multimodal, and context-aware conversational AI.
• Thinking Machines Just Announced More Human Like AI — Thinking Machines Lab introduces "Interaction Models" for real-time, human-like AI collaboration via simultaneous input/output.
• [AINews] Thinking Machines' Native Interaction Models - TML-Interaction-Small 276B-A12B - advances SOTA Realtime Voice and kills standard VAD — Real-time, full-duplex multimodal interaction models are advancing human-AI collaboration beyond turn-based systems.
• Interaction Models: A Scalable Approach to Human-AI Collaboration — Native, real-time multimodal interaction within AI models enhances human-AI collaboration beyond turn-based systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.