Lin Junyang, 31, former technical lead of Alibaba's Qwen large model.

On August 12, he announced in Shanghai the founding of a new company: Pragmatik Labs, abbreviated as p7k. The research direction: "next-generation agents spanning the digital and physical worlds." Investors include Tencent. He left Alibaba in March this year.

The attention this news generated in the AI circle has nothing to do with "another founder quit to start a company" — those stories happen every day. The key question is: who is Lin Junyang, what direction is he pursuing, and why did he choose to make this move in August 2026?

The Qwen Tech Lead's Departure: A Signal Worth Reading

Lin Junyang's role in the Alibaba Qwen team is hard to summarize in a single sentence. Based on public information, he was one of the core technical leads of the Qwen large model, involved in Qwen's journey from inception through continuous iterations. Qwen is Alibaba's most important bet in the general-purpose large model space, and Lin Junyang was one of the technical helmsmen behind it.

In March 2026, he left Alibaba. At the time, the industry broadly expected him to join another major tech company in a technical management role — the most common career path in China's LLM circle. But four months later, in August, he chose a different path: entrepreneurship.

The choice itself is a signal. In the current LLM startup wave, the vast majority of founders combine "algorithm engineer + business partner" — but Lin Junyang's unique edge is that he has not only technical depth but also product-direction judgment. More importantly, his chosen research direction isn't "build another large model" — it's "build what sits above the model": Agent.

p7k's English name, Pragmatik Labs, is noteworthy — "Pragmatic" means practical, pragmatic. This creates an interesting contrast with the current AI industry atmosphere: the past two years of LLM startups were almost entirely a "parameter arms race" — who has the biggest model, the cheapest cost, the fastest inference. p7k's choice goes the opposite direction: skip the parameter war, skip the benchmark chasing, focus on solving "how to actually use models."

Agents Spanning Digital and Physical Worlds: What's Next for the Decade?

Lin Junyang's announced p7k research direction — "next-generation agents spanning the digital and physical worlds" — is an ambitiously positioned goal.

Digital-world agents we've already seen many of: coding agents (Cursor, Devin), office agents (Microsoft Copilot, ByteDance Doubao), customer service agents. Their common feature: they run in purely software environments — inputs are text or code, outputs are text or code.

Physical-world agents are entirely different. Robotics, autonomous vehicles, automated systems in smart factories — these agents perceive physical environments, make real-time decisions, and interact with the physical world. Their inputs are sensor data (vision, force, position); their outputs are mechanical actions.

Lin Junyang's direction is to bridge these two. Imagine a scenario: you tell your agent in the digital world, "help me plan tomorrow's work." The agent doesn't just handle emails and schedule meetings on your computer — it can also command physical-world devices: turn on the office lights, adjust the thermostat, dispatch logistics robots to move cargo. That's the vision of "digital-physical cross-domain agents."

This direction is technically extremely difficult — but also highly valuable. If achieved, it solves the biggest pain point in current AI: AI can help you write code and edit slides on screen, but it can't help you tighten a screw, move a box, or repair a machine. Extending Agent capabilities from the digital world into the physical world is the critical step for AI to evolve from a "content generation tool" into a "general-purpose action tool."

Of course, p7k is still in its very earliest stages. Team size, funding, product form — none have been made public. But Lin Junyang's background, the前瞻性 of his research direction, and Tencent's capital support make this a company worth watching closely.

For Dawn Vision readers, Lin Junyang's entrepreneurial story sends a more important signal: top LLM talent is flowing from the "foundational model layer" toward the "application layer." This isn't the first time — core members of OpenAI, Anthropic, DeepSeek have all gone on to build application-layer products after leaving. But when the Qwen tech lead takes this path too, it signals a trend that has crystallized: the foundational model layer's competitive landscape is gradually taking shape, while the application layer's innovation space remains enormous.

For anyone considering entering the AI space, Lin Junyang's choice may be more instructive than any analysis report: don't just stare at model parameters and benchmarks. The real opportunity lies in "what models can actually do" — and Agent is the answer to that question.

See you tomorrow.