Just now, Lin Junyang officially announced the launch of his startup Pragmatik Labs.

Just now, Lin Junyang, former technical lead of Alibaba Qwen, officially announced his new entrepreneurial move: he has founded a new AI company named Pragmatik Labs in Shanghai, short for p7k. The company's research direction is also officially disclosed for the first time: to develop next-generation Agents that span the digital world and the physical world.

Meanwhile, the investor lineup of Pragmatik Labs has also been unveiled. Lin Junyang revealed that Gaorong Ventures and HSG co-led this round of financing, with support from Tencent and Shanghai Future Industry Fund.

Earlier sources disclosed that the financing size of this round for Pragmatik Labs reached hundreds of millions of US dollars, with Gaorong Ventures and HSG each investing about 100 million US dollars, Tencent investing about 20 million US dollars, and the company's post-money valuation is around 2 billion US dollars.

This means that before its official public debut, the valuation of Pragmatik Labs has already reached the level of tens of billions of RMB.

So far, from leaving Qwen in March this year, to being exposed for starting a business in May, and the financing progress being revealed in June, the nearly half-year speculation around Lin Junyang's next move finally has an answer.

This technical lead, who once led Qwen to enter the first tier of global large models, did not choose to stay in a large tech company, nor did he simply limit the goal of his new company to foundation models.

He bet his next stop on Agents.

At present, the official website of the new company has also been launched, but the information disclosed is very limited, with no model parameters announced and no specific products released. However, the company's technical layout is already quite clear.

Pragmatik clearly divides its research directions into two parts: Digital Agents + Physical Agents, based on which the directions are refined into four sections:

Digital Agents: General-purpose agents oriented to knowledge work, business operations and industrial-level workflows.

Physical Agents: Embodied intelligence that can adapt to the environment, take actions, and complete long-horizon tasks in the real world.

From Research to Product: Implement cutting-edge research into real products, and continuously shape the next research direction through feedback from the real world.

Long-term Exploration: Build systems that break through existing paradigms to accelerate scientific progress.

In fact, this technical route choice is not without omens. As early as March this year, Lin Junyang published the article *From 'Reasoning' Thinking to 'Agentic' Thinking*, proposing that the focus of the next stage of AI will further shift from "training models" to "training agents, and even training systems composed of multiple agents".

Article link: https://justinlin610.github.io/blog/from-reasoning-to-agentic-thinking/

Lin Junyang has gone further in defining "Agent".

In his vision, what really needs to be trained in the Agent era is the complete system composed of "model + environment", including Harness, training environment, evaluator, and even the collaboration mode between multiple Agents.

This is probably the "next-generation Agent" mentioned by Pragmatik Labs.

When talking about the name of the new company, Lin Junyang explained: "I initially studied linguistics because a friend recommended pragmatics to me, and later I turned to computational linguistics and natural language processing (NLP).

This name, on the one hand, means returning to the place where everything really happens; on the other hand, it points to pragmaticism — we believe this is exactly the direction AGI should pursue."

Speaking of this, we have to review Lin Junyang's impressive resume again.

Lin Junyang was born in 1993. He completed his undergraduate study in the English Department of University of International Relations, and learned Japanese, Russian, German and French at the same time; he completed his master's degree in Foreign Linguistics and Applied Linguistics at Peking University. His profound linguistic background laid the foundation for him to understand the underlying logic of large language models in the future, and gave him more in-depth cognition and accumulation in the direction of machines understanding human language and intentions.

According to public information, during his university years, Lin Junyang used an English-German vocabulary database containing more than 3 million samples for testing, and found that the accuracy of normal machine translation was only about 23%, while the early version of Transformer could reach 27%. This seemingly insignificant data comparison made him keenly capture the future explosive potential of large models.

After graduating with a master's degree in 2019, Lin Junyang joined the Intelligent Computing Laboratory of Alibaba Damo Academy as a senior algorithm engineer, engaged in natural language processing research, and officially entered the cutting-edge field of AI.

After that, in just over 6 years, he started his path of rapid career promotion...

At the end of 2022, Alibaba merged the AI teams such as the language and vision teams of Damo Academy into Alibaba Cloud, and established the Tongyi Lab. Lin Junyang officially took over the Tongyi Qwen series of large models and served as the technical lead.

Under his leadership, Alibaba launched the Qwen series of open-source models covering various parameter scales, which truly pushed Alibaba's AI technology into the broader industry ecosystem.

In August 2024, after Zhou Chang, the former head of Tongyi Qwen, left for ByteDance, Alibaba carried out a general salary adjustment for employees in core departments, and Lin Junyang was promoted to P9 as a result.

In May 2025, with the excellent achievements of his team, Lin Junyang was promoted to P10, becoming the youngest P10-level technical lead in Alibaba.

In October 2025, Lin Junyang personally set up a robotics and embodied intelligence team inside Qwen, which was regarded by the outside world as one of Alibaba's most clear exploration strategies for physical AI systems. At that time, he posted on X that he was trying to let the model step out of the virtual screen, truly understand the space and learn to operate physically, so as to push intelligence to the real physical world.

Shortly after, in early March this year, Lin Junyang posted a tweet announcing that he was leaving the Qwen team.

And just now, Lin Junyang officially announced his new entrepreneurial direction — Pragmatik Labs, choosing to bet on "next-generation Agents that span the digital world and the physical world".

Lin Junyang once participated in building one of the most globally influential large model systems in China. Now, the fact that he wrote both the digital world and the physical world into the company's research goals at the first time of announcing the new company's name is itself a noteworthy industry signal.

The "war" of foundation models is far from over, but more and more top researchers have begun to ask the next question: Models have been able to think, so how do we get them to actually get things done next?

What Pragmatik Labs is going to answer may be exactly this question.

Reference Links:

https://x.com/JustinLin610/status/2087214239643730227

https://pragmatik.com/

This article is from the WeChat official account "Almosthuman" (ID: almosthuman2014), author: Almosthuman focusing on Agents, editor: Youli, Shanhui, published with authorization from 36Kr.