Thinking Machines Lab
E246｜何谓蒸馏？聊聊硅谷如何看中国开放模型逼近前沿 adds Thinking Machines Lab to the open-model strategy branch. The source says its open-weight model release and Tinker-style post-training service show that U.S. labs may also use openness to create demand for training, adaptation, or downstream services rather than only to give away a model.
Thinking Machines Lab is the model lab named in 171: 【AI季报 26Q2】从 coding 到 RSI，强者愈强的未来？ through its release of Interaction Model. The source says the model is a 276B-parameter MoE with 12B active parameters, trained from scratch for real-time multimodal interaction.
In the episode, Thinking Machines Lab matters because it pushes Voice Interaction beyond turn-taking chat. The company is used as the contrast to “walkie-talkie” voice assistants: the source argues that AI interaction may need to listen, speak, see action, and be interrupted like a phone call rather than a text chat with audio attached.
Connections
- Open Source AI Models, Open-Weight Commercial Licensing, and Model Sovereignty / 模型主权 — E246’s open-model and post-training-service strategy branch.
- Interaction Model — model/product attributed to Thinking Machines Lab in the source.
- Voice Interaction — interaction category it extends.
- OpenAI — comparison point through OpenAI real-time voice products.
- AI Commercialization Pressure — model labs still need product surfaces where capability becomes user value.