Qwen Founder Junyang Lin Launches Pragmatik Labs for Digital and Physical AI Agents

Just five months after stepping away from Alibaba’s Qwen — the most-downloaded open-source large language model family on Earth — Junyang Lin has officially pulled back the curtain on his next act. On August 12, 2026, Lin announced the founding of Pragmatik Labs (also stylized as p7k), a Shanghai-based startup dedicated to building next-generation AI agents that span both digital knowledge work and physical, embodied intelligence.

The announcement, teased for weeks across Chinese tech media and confirmed by Lin on his personal X account (@JustinLin610), marks one of the most closely watched founder launches in the global AI space this year. Lin is not just another researcher going independent. As the technical lead behind Qwen, he oversaw the model family that, by early 2026, had surpassed Meta’s Llama to become the single most-downloaded LLM lineage on Hugging Face — accumulating over 700 million downloads and, by some tallies, accounting for more than 18 percent of all top-1,000 model traffic on the platform.

From Qwen to Pragmatik

Lin’s departure from Alibaba on March 3, 2026, sent shockwaves through the open-source AI community. Qwen had become the crown jewel of Chinese open-source AI, a project that Alibaba Cloud had poured enormous resources into across text, vision, code, and multimodal variants. Lin had been the public face and technical architect of that effort, presenting at major conferences and driving the roadmap that pushed Qwen from a regional contender to a globally dominant open model family.

After leaving, Lin spent several months publishing as an independent researcher. In a widely circulated July 2026 talk, he argued that the field was undergoing a fundamental shift — away from training ever-larger static models and toward building agents that can plan, act, and learn over long time horizons. He was blunt about what he saw as the limitations of the “hybrid thinking” paradigm that many frontier labs, including his former team, had been pursuing. That thesis now serves as the intellectual foundation for Pragmatik Labs.

Four arenas of focus

According to Lin’s announcement and the company’s newly launched website at pragmatik.com, Pragmatik Labs is organized around a dual-realm thesis: the next generation of AI must operate fluently across both digital and physical environments. The startup describes its mission as building “agents for digital and physical worlds, built to transform work and life.”

The company’s focus reportedly spans four interconnected arenas:

1. Digital agents — AI systems that transform knowledge work, operating software, managing workflows, and completing complex multi-step tasks in digital environments.
2. Physical and embodied agents — embodied intelligence that adapts to and acts within the real physical world, solving long-horizon tasks in unstructured environments.
3. Foundation models for agency — the underlying model research needed to support robust planning, sample-efficient interaction, and reliable tool use.
4. World-modeling infrastructure — the systems and benchmarks needed to train and evaluate agents that must understand physical dynamics and consequences.

This is an ambitious scope. While dozens of startups are pursuing digital agents — the current darlings of Silicon Valley — far fewer are attempting to unify digital and physical intelligence under one roof. Lin’s bet is that the same foundational capabilities, particularly long-horizon planning and sample-efficient learning, are shared across both domains and that building them together creates compounding advantages.

The embodied AI focus is not entirely new for Lin. In October 2025, while still at Alibaba, he personally established a robotics and embodied intelligence team inside Qwen — a move that outside observers at the time read as a signal of where his interests were heading.

A $2 billion bet backed by China’s top investors

Pragmatik Labs has assembled a formidable cap table. The company’s funding is led by Sequoia China and Gaorong Capital, with participation from Tencent and the Shanghai Future Industry Fund. According to earlier reporting from The Information and 36Kr, Tencent committed approximately $20 million in the first round, bringing the total initial raise to several hundred million dollars. The post-money valuation is reported to be around $2 billion, or roughly 13.5 billion yuan.

That a Tencent-led syndicate is backing the former face of a rival’s flagship AI project is itself a notable signal. Tencent has its own large model efforts — including the Hunyuan family — but the investment suggests a strategic belief that the agent layer, and Lin’s particular approach to it, represents a distinct and high-value opportunity that complements rather than competes with foundation model development.

The involvement of the Shanghai Future Industry Fund also places Pragmatik Labs squarely within the municipal government’s push to build Shanghai into a global hub for embodied AI and next-generation robotics. China’s national strategy has increasingly funneled resources toward physical AI, with both central and local governments viewing embodied intelligence as a domain where the country can establish durable advantages.

The broader agent wave

Pragmatik Labs does not launch in isolation. Lin’s return coincides with a broader wave of agent-focused startups emerging from China. Manus, the high-profile agent startup that captured attention earlier in 2026, has also been making moves — and observers have noted that Lin and the Manus team are, in the words of one industry analysis, “pointing in the same direction.” Both are betting that agents, not raw model scale, will define the next phase of AI commercialization.

Globally, the shift is unmistakable. NVIDIA CEO Jensen Huang has repeatedly framed the trajectory from generative AI to agentic AI to physical AI as the defining roadmap of the decade. Capgemini’s 2026 research institute study of nearly 1,700 executives found that physical AI adoption is accelerating rapidly across industries. And on the technical side, research on world models, long-horizon planning, and sample-efficient reinforcement learning — the exact problems Pragmatik Labs is targeting — has become one of the most active frontiers in the field.

What to watch

Pragmatik Labs is at day one. It has no shipping products, no public benchmarks, and no disclosed technical details beyond the high-level mission statement. What it does have is arguably the most proven open-source model builder in the world, a war chest of several hundred million dollars, backing from China’s most connected investors, and a clear and contrarian thesis about where AI is heading.

The key questions for the coming months: How quickly can Lin assemble a world-class team in an intensely competitive talent market? Will the dual digital-physical approach prove to be a genuine technical advantage or an unfocused dilution of effort? And perhaps most importantly — can Pragmatik Labs produce agent capabilities that meaningfully surpass what incumbent labs with far more compute can achieve?

Lin’s track record with Qwen suggests betting against him would be unwise. The open-source community will be watching closely.