Qwen chief's Pragmatik Labs raises $220 million at $2 billion valuation

Five months after leaving Alibaba Group Holding Ltd., the architect of its Qwen large language model has founded Pragmatik Labs in Shanghai to build AI agents spanning digital and physical work, backed by a $220 million angel round at a $2 billion valuation.

In a March essay titled "From 'Reasoning-Based Thinking' to 'Agent-Based Thinking,'" Lin argued that reasoning models seek optimal answers in closed contexts while agents must think to act, revising plans through interaction with an uncertain environment — the intellectual foundation for the new company.

The round was co-led by HongShan, formerly Sequoia China, and Gaorong Ventures, each investing about $100 million, with Tencent Holdings Ltd. contributing roughly $20 million and the Shanghai Future Industry Fund also participating, people familiar with the matter said. The company, internally abbreviated as p7k, will develop digital agents for knowledge work and embodied AI for physical environments.

The valuation — roughly 14 billion yuan — prices in Lin's record building Qwen from zero into one of the most widely used open-source model families, and signals where Chinese AI capital is heading: away from parameter scaling and toward agents that operate browsers, code repositories, and eventually robots.

The pivot was foreshadowed months ago. After leaving Qwen in March — days after Elon Musk publicly praised the model's "impressive intelligence density" — Lin posted a minimalist farewell on X: "me stepping down. bye my beloved qwen." Alibaba Group CEO Wu Yongming confirmed the departure in an internal email, reaffirming the company's open-source strategy.

Pragmatik's website lists two business lines. Digital agents target knowledge work, enterprise operations, and industrial processes, requiring autonomous reasoning, tool invocation, and long-horizon task coordination across days or weeks. Physical agents extend that capability into real environments, handling tasks that take tens of minutes to hours. The name "Pragmatik" reflects both Lin's academic roots in pragmatics — the linguistics field he studied before moving into computational linguistics and large models — and his view that artificial general intelligence must be pragmatic, solving problems rather than merely predicting text.

Two environments, one bottleneck

Placing digital and physical agents in one company is Pragmatik's most ambitious decision. The engineering gap is wide: digital agents operate on browsers and code repositories with near-zero trial-and-error costs, while physical agents confront robot hardware, sensors, and physical laws where a single mistake can damage equipment. Most startups would pick a single point of entry. Lin's framework treats digital and physical as two environments facing the same core bottleneck — maintaining goals over long time horizons, handling uncertainty, and learning from environmental feedback. The digital world serves as a training ground for general capabilities like long-horizon planning and memory, which can transfer to physical systems; physical demands for causal reasoning feed back to sharpen the digital side.

Capital structure maps the route to market

The shareholder lineup sketches a commercialization path. Gaorong and HongShan bring capital and talent pull; Tencent and the Shanghai Future Industry Fund complete the distribution and application-scenario pieces. If Pragmatik first breaks through in digital agents, Tencent's WeChat, WeCom, Tencent Docs, and cloud infrastructure become natural deployment channels. The Shanghai fund anchors the physical-agent half: the city holds China's most concentrated robotics cluster, industrial manufacturing base, and chip supply chain — a reason Lin chose Shanghai over Beijing or Hangzhou.

The $2 billion valuation, set before any public product or revenue, is a bet on Lin and the core team he brought from Qwen. For Alibaba, the departure removes a technical leader who built its flagship open-source franchise; for Tencent, the roughly $20 million stake buys a window into the agent race without building a rival foundation model. The broader shift: Chinese AI capital is rotating from the parameter-scaling arms race toward agent systems that close the loop from model to action — testing whether the industry's next revenue pool is software that finishes work, not models that chat.

This article is for informational purposes only and does not constitute investment advice.