NewsFlash Articles Data Fundraising BlockBeats Pro

Manus and Lin Junyang, Collective Comeback

Read this article in 15 Minutes

These two returns point to the same shift: AI competition is now moving from the model's capabilities per se to how the model accesses tools, adapts to the environment, and accomplishes real-world tasks.

Original Title: "Manus and Lin Junyang, Both Make a Comeback"

On August 12, two star projects and figures that had been highly regarded in the AI community announced their comeback.

On August 11, local time in the United States, AI startup Manus issued a notification to users, stating that the company would "soon resume operation as an independent entity." As part of the separation from Meta, data generated by some users on and after December 29, 2025, will be deleted in accordance with regulatory requirements.

According to the company's notification, affected users need to complete a data backup by 7:59 AM Singapore time on August 23. During August 23 to August 24, the relevant data will be deleted, and the affected accounts will temporarily be inaccessible. Starting from 8:00 AM on August 25 Singapore time, users can restore the previously backed-up data and resume using the service.

Manus stated that this adjustment was not due to a data breach or security incident but was a measure taken to meet specific jurisdictional regulatory requirements during the process of the company's return to independent operation. Unaffected users do not need to take any action and can continue to use the service as usual.

Several hours later, another important comeback announcement came from Lin Junyang. The former Alibaba Qwen Project Technical Lead announced the establishment of the AI lab Pragmatik(p7k) Labs in Shanghai, focusing on research on the next generation of intelligent entities in the digital and physical worlds.

Both are projects involving intelligent entities, one related to a company's return and the other to a technical lead's new venture, but both point to the same shift: AI competition is transitioning from the capabilities of models themselves to how models invoke tools, adapt to environments, and complete real-world tasks.

Meta Acquisition Transaction Moves to Unwinding Phase

The Manus-Meta transaction began on December 29, 2025.

On that day, Meta announced its plan to acquire Manus. The company was founded in 2022, initially developed in China, and later relocated to Singapore, focusing on AI intelligent agent products.

At the time, both parties did not disclose the transaction amount, but previous reports indicated that the deal was valued at around $2 billion, with a potential total value of up to $2.5 billion, including employee retention arrangements.

After the acquisition, Meta plans to apply Manus' AI assistant technology to its own consumer and enterprise products to enhance its AI capabilities.

However, this transaction later came under regulatory scrutiny.

On April 28, CCTV News reported that the much-watched Manus acquisition case had come to a conclusion. The Office of the National Foreign Investment Security Review Working Mechanism (National Development and Reform Commission) made a decision to prohibit foreign investment in the Manus acquisition project in accordance with the law and regulations, requiring the parties involved to cancel the acquisition transaction.

Subsequently, Manus and Meta began to proceed with operational separation and ceased data sharing between the two parties.

According to previous information, the operational separation between the two parties was completed in May. Currently, Manus is in the final stages of data processing and preparing for independent operation.

This data adjustment mainly involves data generated on and after December 29, 2025.

Manus stated that the related data deletion was to meet regulatory requirements. The company has launched data backup and recovery tools to help affected users preserve task records.

Affected users can back up their data multiple times during the backup window. If users continue to generate new task data after the initial backup, they need to back up again to ensure the latest data is saved.

Manus mentioned that users will not be charged during the backup period. After restoring the account, the company will also provide a return reward.

For users who registered using an Apple ID or Facebook account and may not have a corresponding email address, Manus reminded them to pay attention to in-app notifications.

Manus Founding Team and Product Roadmap

The company behind Manus is Butterfly Effect, founded in 2022.

Founder Xiao Hong had a long history in enterprise software development. Records show that he studied software engineering at Huazhong University of Science and Technology. After graduating in 2015, he founded Wuhan Nightingale Technology, under which the WeChat tools "Yiban Assistant" and "Weiban Assistant" served over 2 million business users.

Butterfly Effect later launched Monica, an AI assistant product that integrates multiple language model capabilities. Manus further expanded AI capabilities into the task execution domain.

Manus is positioned as a general AI agent platform, where users can use natural language commands to instruct the system to perform tasks such as information retrieval, report writing, document processing, and code development.

Co-founder and Chief Scientist Ji Yichao had previously developed an iPhone browser and founded Peak Labs, focusing on information extraction and search technology. Co-founder Zhang Tao had previously worked on product-related tasks at ByteDance and LightYear.

On March 6, 2025, Manus was officially launched to the public.

Rapid Growth Pre-Acquisition, Future Ownership Still Undecided

Prior to being acquired by Meta, Manus had completed multiple funding rounds and experienced rapid growth.

Earlier in 2025, the company raised $75 million from Benchmark, with a post-money valuation reportedly around $5 billion.

In December 2025, Manus announced that eight months after product launch, their annualized recurring revenue had exceeded $1 billion, with a total annual revenue run rate of $1.25 billion.

The company also disclosed that the platform had processed over 147 trillion tokens and created over 80 million virtual machines.

At the time of Meta's acquisition of Manus, both parties aimed to further expand the AI agent business. However, following the completion of the transaction, Manus's future ownership structure and funding arrangements have yet to be disclosed.

According to previous reports, during the transaction completion, the founding team had explored raising approximately $1 billion to buy back company shares.

Some early investors were also involved in the discussions, including firms like ZhenFund and IDG Capital.

Post-acquisition, Manus stated that once it resumes independent operations, the company will continue to serve global users and plan to introduce new product features. There are plans to further enhance AI agent capabilities in the future, although no specific product roadmap has been announced.

Regarding data storage, Manus mentioned that after resuming independent operations, user data will continue to be stored in the U.S. and Singapore.

From Training Models to Building Agents, Lin Junyang Embarks on a New Exploration

With Manus returning to independent operations, this startup, with a focus on general agents, will once again have the space to independently adjust its product and funding arrangements. On the same day, Lin Junyang, another key tech leader who had previously led domestic large-scale model development, also shifted his entrepreneurial focus to agents.

Lin Junyang announced his resignation from Alibaba on March 3, 2026. Previously, he was deeply involved in large-scale model development, focusing on inference capabilities, intelligent agent training, and the interaction of models with the environment during the development of the Qwen series of models.

The shift from inference models to intelligent agent systems signifies a changing focus in AI competition. Future technological breakthroughs will stem not only from the models themselves but also from how models integrate with tools, environments, and real-world tasks.

Lin Junyang's new research direction is centered around intelligent agent systems. The name Pragmatik Labs is derived from "Pragmatics" in linguistics. Lin Junyang explained that he initially studied linguistics because a friend recommended pragmatics to him, and he later transitioned to computational linguistics and natural language processing. "Pragmatik" represents a return to where things happened, while also reflecting the team's understanding of pragmatism, namely that AGI ultimately needs to address real-world practical issues.

Pragmatik Labs will focus its research on two directions: digital agents and physical agents.

In the digital realm, the company aims to build general agents for knowledge work, enterprise operations, and industry processes, enabling AI to handle more complex tasks. In the physical realm, the company is exploring embodied intelligent systems that can enter real-world environments, adapt to changes, and perform long-term tasks.

Pragmatik Labs introduces on its official website that the next-generation agents need to possess the ability to reason, invoke tools, learn from feedback, and coordinate actions. Unlike traditional models that mainly focus on information generation, intelligent agents need to operate continuously in the environment, adjust strategies based on outcomes, and achieve long-term goals.

This direction aligns with Lin Junyang's previous proposition of "training models → training intelligent agents." In his view, inference models address how to make models more effectively compute internally before responding, while intelligent agent systems focus on how to make models take action in a real environment. Future AI systems will need to address issues such as tool selection, task planning, environment feedback, failure correction, and multi-round task collaboration.

Pragmatik Labs is currently laying out its strategy in this direction, aiming to extend AI capabilities from digital information processing to real-world tasks.