Thinking Machines Lab unveils full-duplex AI vision, three months after Chinese rival open-sourced similar tech
The AMW Read
The article introduces a new entrant (TML) to segment 01 and details a novel full-duplex architecture, but the core concept was already published by a Chinese lab, making this an incremental confirmation rather than a breakthrough.
Thinking Machines Lab unveils full-duplex AI vision, three months after Chinese rival open-sourced similar tech
Thinking Machines Lab (TML), the startup founded by former OpenAI CTO Lilian Weng, published its first technical vision today: a full-duplex real-time conversational AI model that can process and generate multiple modalities simultaneously. The company released a demo video showing the model engaging in seamless, interruptible dialogue. However, the article notes that Chinese firm Shengwu Intelligence (面壁智能) released and open-sourced its MiniCPM-o 4.5 model with the same core capability—an 'Omni-Flow' framework based on time-aligned micro-turns—three months earlier in February 2026.
Why it matters: This is a textbook 'concept-engineering moat' pattern applied to the emerging full-duplex AI interaction paradigm. TML's announcement validates the technical direction pioneered by a lesser-known Chinese lab, confirming that the next frontier in human-AI interaction is the transition from turn-based (VAD) to continuous, streaming multimodal dialogue. The three-month lead and open-source delivery by Shengwu Intelligence undercuts TML's narrative leap, forcing the question of whether Western labs can sustain a first-mover advantage when Chinese competitors are willing to fully open their code. For enterprise buyers evaluating voice AI platforms, the existence of an open full-duplex baseline compresses moats built on closed demos.
Grounded expert take: The convergence on a nearly identical architecture—time-aligned micro-turn processing of audio, video, and text tokens—between two independent teams is a strong signal that this is the dominant design for real-time multimodal AI. Shengwu's earlier delivery with an Apache 2.0 license turns this into a democratic capability, not a proprietary differentiator. The article positions Shengwu as the 'true first mover,' and TML as a high-profile verifier—a dynamic that benefits the open-source ecosystem more than any single lab. The real competition now shifts from who shows the demo to who can scale reliable, low-latency inference at the edge and integrate into devices.
#FullDuplexAI #MultimodalInteraction #OpenSourceAI #ChineseAI #RealTimeVoice #ThinkingMachinesLab
