The "Lin Junyang Phenomenon" in the AI venture capital circle
As expected, Lin Junyang, who has not appeared in public for a long time, announced the official debut of his new company PragmatikLabs.
Although there are no product demos or release schedules yet, this Shanghai-registered company has been jointly led by Gaorong Capital and HSG, with Tencent and Shanghai Future Industrial Fund participating in the follow-on investment. Its valuation at the angel round has hit 2 billion US dollars.
According to the industrial and commercial structure of the domestic operating entity Yuyong (Shanghai) Technology, the three external institutions have invested a total of about 220 million US dollars, only taking 12% of the equity. The founder Lin Junyang and related entities hold 88% of the shares.
The information revealed by this equity structure may be richer than the financing amount. The founder and the team firmly hold the control, the financial investors hold limited shares, and each strategic investor takes a seat. From this perspective, it is more like capital scrambling for admission tickets to a scarce target, exchanging a high price for a small number of equity, which is a typical transaction structure of "talent premium".
But the 2 billion US dollars is only the price offered by capital in advance. A person familiar with the matter disclosed that the financing has just completed the delivery process, and the team has begun to seek the next round of financing. Another source said that the new round of valuation may have climbed to 5 billion US dollars. The product has not yet been unveiled, but the valuation has continued to rise on the way.
01
The Operator of the Open Source Ecosystem
Lin Junyang, born in 1993, took an atypical path. He studied in the English Department of University of International Relations for his bachelor's degree, and was admitted to the School of Foreign Languages of Peking University for his master's degree to study foreign linguistics and applied linguistics. During his master's period, he turned to computational linguistics, thus entering the NLP field. He is proficient in six languages, and this humanistic background later profoundly influenced his understanding of large models.
After graduating in 2019, Lin Junyang joined Alibaba DAMO Academy through campus recruitment, and worked on pre-trained model R&D in the M6 team led by Yang Hongxia. At the end of 2022, Alibaba adjusted its organization to establish the Tongyi Lab, and Lin Junyang was appointed as the technical lead of Qwen. He was 29 years old that year.
The subsequent story has been told repeatedly.
In August 2023, Qwen was first open-sourced, and then released models at an almost aggressive pace: covering all sizes from 0.5B to 72B, from language models to Coder, Math, visual multimodal, and then to the reasoning model QwQ. In the Qwen3 era, Qwen has formed a complete product matrix.
By the beginning of this year, the number of Qwen-derived models has exceeded 200,000, with more than 1 billion global downloads. Alibaba has released more than 400 Qwen open-source models since 2023. According to Hugging Face's 2026 Spring Report, downloads of Chinese open-source models have accounted for 41% of the global total, and Qwen is one of the most important contributors. Together with DeepSeek, it forms the two pillars of China's open-source large models.
At that time, Lin Junyang played a far more important role than just the technical lead. He was the main promoter of the open source strategy and almost the spokesperson of Qwen in the global developer community. When developers asked whether QwQ-Max would be open, he directly replied "we will opensource the models".
A person close to the Qwen team recalled that Lin Junyang's management style is to set clear "targets" and let the team iterate quickly. He does not pursue perfection in every detail, but pays more attention to whether things can be launched first. He believes that the lead does not need to be proficient in every line of code, but must understand the underlying "physical logic", know why it works in this way, and know how to grasp the big picture and let go of minor issues.
With far fewer resources than its competitors, Qwen chose a differentiated path: using a full-size, full-modality open source matrix to cover the widest range of developer groups. Small models can run on mobile phones or even microcontrollers, and large models are benchmarked against GPT-4, so developers can always find the one that suits them.
But this path hit an inflection point in early 2026.
At the end of February, Alibaba adjusted the architecture of its large model business, and its strategic focus shifted to commercialization. According to multiple media reports, Lin Junyang's proposition of "extreme open source and zero commercial cost" had a structural conflict with the group's commercialization demands. After an internal meeting in March, Lin Junyang left the meeting room and submitted his resignation. Yu Bowen, the post-training lead, and Li Kaixin, the core contributor, later confirmed their departure.
At that time, Lin Junyang wrote on his WeChat Moments: "I didn't know so many people in the world love me until these days." He said that at least in his heart, he felt that he had achieved "doing good for the brothers, doing good for Alibaba Cloud, and doing good for the group".
The essence of this resignation turmoil is the collision of technical idealism and the commercial logic of large companies. Alibaba needs returns from AI investment, and the 380 billion yuan investment cannot only bring community reputation; Lin Junyang believes that the core competitiveness of Qwen comes precisely from the open source ecosystem, and excessive commercialization will hurt the trust of developers.
There is no right or wrong in the choices of the two, but one thing is certain: the number of people who have fully operated a global-level open source model family and experienced pre-training, post-training, multimodality, open source ecosystem and large-scale engineering is very limited across the entire industry.
Models can be retrained, computing power can be purchased, and data can be accumulated. But the ecological experience of leading 200,000 derivative models and 1 billion downloads from 0 to 1 cannot be achieved quickly.
This is what capital saw on that night in March.
02
The Multiplier Effect of Talents
When Lin Junyang just left his job, the reaction of the investment circle was very direct. It is said that some investors asked people around to "seek connection", and some people said that "no matter whether he starts a business or not, grab him first". Zhuang Minghao, who focuses on AI investment, judged at that time that if talents of this level start a business, "they will be snatched up in the first round".
Nie Wei, a senior headhunter in the AI field, once used the "multiplier effect" to explain the high premium of top AI talents. When enterprises introduce top talents with high salaries, what they buy is far more than the direct output of one person. "A talent with an annual salary of tens of millions of yuan may create greater benefits than a team of more than ten people."
Top talents themselves are magnets, which can attract outstanding talents in the same field to gather continuously. "The talent pool is so small. If you don't grab them, others will."
This scarcity is further amplified in the AI startup market. According to PitchBook's official Q1 AI VC Trends report, in the first quarter of 2026, global AI startups raised a total of 2555 billion US dollars, exceeding the sum of the whole year of 2025. The valuation of Moonshot AI rose from 4.3 billion US dollars to 31.5 billion US dollars within half a year, and the negotiation price of DeepSeek in the primary market is about 45 billion US dollars.
In the foreign AI circle, Thinking Machines Lab completed a 2 billion US dollar financing when its product was also in the early stage, with a valuation of 12 billion US dollars. Later, it signed an agreement with NVIDIA to deploy at least 1 GW of Vera Rubin systems, with a total investment of up to 600 billion US dollars. Essential AI raised financing at a valuation of 8.6 billion US dollars without revenue, and the logic of investors' betting also revolves around scarcity.
In this case, the traditional DCF valuation model completely fails in these cases. The value of a pre-revenue AI company depends on three variables: the credibility of the founder's resume, the team's judgment on the technical paradigm, and the strategic investors' competition for ecological niches.
Lin Junyang's resume just meets these three variables at the same time. The success of Qwen proves that he can lead the team to win big battles, the long article he published in March proves that he has a clear judgment on the next stage, and the entry of funds has a clear strategic intention.
Tencent has continuously invested in Zhipu AI, MiniMax, and Moonshot AI in recent years. This bet on Pragmatik continues the strategy of locking external cutting-edge AI capabilities through capital. Some analysts explained this: if Pragmatik first makes a breakthrough in the digital Agent direction, the ecosystem composed of WeChat, WeCom, Tencent Docs and Tencent Meeting will provide a natural test field.
All four external shareholders have invested in Moonshot AI. This detail implies that capital has formed a consensus on the pricing of AI talents: the credit of top technical leads can be migrated across projects, and institutions that have invested in Yang Zhilin will not miss the next Lin Junyang.
Ten days before the official announcement, Lin Junyang posted four words on X: "Moderate profit". These four words are like a response to his experience in Alibaba, and also set the commercial tone for the new company:
We will not be open source fundamentalists who lose money for reputation, nor will we be a profit-seeking commercial company. AGI should ultimately be pragmatic, truly solve problems and create value, and profits should be moderate.
The name Pragmatik itself is a manifesto. It comes from "Pragmatics" in linguistics, which studies how language generates real meaning in specific contexts.
Lin Junyang explained that he chose to study linguistics back then because a friend recommended pragmatics. Later, he went all the way to computational linguistics, NLP and large models, and came back to this name after going around. At the same time, it also represents Pragmaticism. This name condenses his whole path from academia to industry.
03
Where is the Barrier of Agents
Capital pays for Lin Junyang's past, and the future of Pragmatik Labs is bet on one judgment: AI is moving from the stage of training models to the stage of training Agents.
This judgment has been fully elaborated in Lin Junyang's article From "Reasoning" Thinking to "Agentic" Thinking.
His core point is: reasoning models look for optimal solutions in a closed context, while Agents must think for action.
The conditions for solving math problems are static, and the model can exchange longer thinking chains for higher accuracy; tasks in the real world are dynamic and open, information will be missing, tools will report errors, web pages will be revised, and an early small error may be infinitely amplified after dozens of steps.
"Agentic thinking is reasoning through action," he wrote.
There is a key word in the article: Harness Engineering. Harness originally refers to the infrastructure required to run tests in software testing. Lin Junyang uses it to describe the core capabilities of the Agent era: the tools, memory, environment, permissions, verification mechanisms around the model, and multi-Agent collaboration paradigms will determine the final effect together with the model itself.
The barrier of the next generation of AI companies is shifting from model parameters to the entire system composed of the model and the environment.
And this judgment is being verified by industry data. Some time ago, XLANG Lab released the OSWorld 2.0 benchmark test, expanding computer operation tasks to 108 real long-range workflows. The results show that the best-performing Claude Opus 4.8 can only achieve 20.6% under the strictest full-task completion indicator, and GPT-5.5 is about 13%.
For the bottleneck of Agents, researchers have summarized five types of typical failures: information omission and tracking failure, misalignment of perception and action timing, lack of domain knowledge and workflow, insufficient verification and reflection, and long-range state drift.
The model will forget the initial instruction constraints, deviate from the target after dozens of operations, and cannot repair itself when exceptions occur. This is exactly the problem corresponding to the four verbs on the Pragmatik official website. Reasoning solves "what to think", using tools solves "what to do", learning from feedback solves "how to improve", and coordinating actions solves "how to complete long tasks".
What is really intriguing is Pragmatik's decision to lay out both Digital Agents and Physical Agents at the same time.
Digital Agents take browsers, code libraries and SaaS as the stage, the data is naturally digital, the cost of trial and error is low, and the iteration is calculated on a daily basis. Physical Agents have to face robot bodies, sensors, complex control systems and uncertain physical laws. A single error may damage the hardware, with a long iteration cycle and huge capital expenditure.
Most startups will choose a single point breakthrough. Lin Junyang chose to promote the two lines in parallel. In his technical framework, the digital and physical are only two environmental forms that Agents face.
The Agent operating the computer needs to perceive the screen, call software, and adjust the plan according to feedback; the Agent controlling the robot needs to perceive the space, manipulate the robotic arm, and correct the action according to sensor data. The input and output are different, but the core bottleneck is the same: how to maintain the target over a long span of time, deal with uncertainty, and continuously learn from environmental feedback.
However, ambition and risk are always two sides of the same coin. The 2 billion US dollar valuation does not only correspond to an Agent application or a robotics company. It must prove that it has the opportunity to become the next-generation basic platform.
Digital Agents have to compete with players such as Manus, Devin, and OpenAI Operator, while Physical Agents have to face leading players such as Tesla Optimus, Figure, and Agility. From a technical lead in a large factory to an entrepreneur, the test Lin Junyang faces has also extended from model R&D to team, product and commercialization.
The success of Qwen has Alibaba's computing power, data and brand endorsement, while Pragmatik needs to rebuild all these in an independent environment.
But he has done one thing right: at the node where model capabilities are gradually commoditized, he did not choose to build another "Tongyi", but led the battlefield to systems and environments. This is in the same line with the logic of him promoting Qwen open source back then: when everyone is stacking parameters, find the next dimension that really determines the outcome.
The value migration in the AI industry always happens in an unexpected way. In 2023, people compete for model parameters, in 2024 they compete for the length of the reasoning chain, in 2025 they compete for the open source ecosystem, and in 2026 the focus shifts to whose Agent can really get the work done. Every paradigm shift will redistribute the industry's right to speak, and redefine what assets are the most valuable.
Lin Junyang is precisely the projection of this value migration in the market. As computing power and data become standard configurations and model capabilities tend to be homogeneous, what is really scarce is the kind of person who understands both technology and ecology, has done research and led large-scale engineering, has a clear judgment on the next paradigm, and can form a team to execute.
Capital prices for them in advance, because they know that missing these few people is very likely to mean missing the admission ticket to the next era.
This article is from the WeChat official account "Xinmou" (ID: xinmouls), Author: Lu Yao, published with authorization from 36Kr.