The Thinking Machines, led by former OpenAI CTO Mira Murati and with a valuation skyrocketing to 12 billion yuan, wiped out all the "real-time voice AI" underwear on the market today. They released the first "interaction model" TML Interaction Small, which directly pierced through the sluggishness of traditional AI's "wait until you finish speaking before replying". The response delay is reduced to 0.4 seconds, allowing you to make eye contact while conducting complex searches in the background. Those middleware companies that rely on outsourcing speech interruption and adding visual shells to large language models have been directly cleared. Take a look at the sense of pressure in this wave of action: 2023: The big language model uses text dialogue boxes to allow ordinary people to touch the threshold of AI. 2024: Voice APIs and shell plugins are rampant, and AI can barely listen and speak, but it relies entirely on external code. 2026: Thinking Machines writes full duplex collaboration into their bones, with 200 millisecond micro round processing, completely eliminating the grass platform team of the splicing era.