Alibaba’s Qwen crew of AI researchers have been among the many most prolific and well-regarded by worldwide machine studying neighborhood — transport dozens of highly effective generalized and specialised generative fashions beginning final summer time, most of them completely open supply and free.
However now, simply 24 hours after transport the open supply Qwen3.5 small mannequin sequence—a launch that drew public reward from Elon Musk for its “spectacular intelligence density”—the mission’s technical architect and a number of other different Qwen crew members have exited the corporate underneath unclear circumstances, elevating questions and issues from all over the world in regards to the future route of the Qwen crew and its concentrate on open supply.
The departure of Junyang “Justin” Lin, the technical lead who steered Qwen from a nascent lab mission to a worldwide powerhouse with over 600 million downloads, alongside two fellow colleagues — workers analysis scientist Binyuan Hui and intern Kaixin Li — marks a unstable inflection level for Alibaba Cloud and its position as a global open supply AI chief.
These three Qwen Staff members introduced their departures on X right this moment, although they didn’t share the explanations or whether or not or not it they had been voluntary. VentureBeat reached out to sources at Alibaba for extra info and can replace once we acquire it. Lin himself signed off with a easy put up: “me stepping down. bye my beloved qwen.”
Whereas the corporate celebrates a technical triumph, the sudden exit of its core management suggests a deepening rift between the researchers who constructed the fashions and a company hierarchy now pivoting towards aggressive monetization.
The departing researchers’ remaining reward: pocket-sized intelligence
The Qwen3.5 small mannequin sequence (starting from 0.8B to 9B parameters) represents a remaining masterstroke in “intelligence density” from the founding crew.
The fashions make use of a Gated DeltaNet hybrid structure that enables a 9B-parameter mannequin to rival the reasoning capabilities of a lot bigger methods.
By using a 3:1 ratio of linear consideration to full consideration, the fashions keep a large 262,000-token context window whereas remaining environment friendly sufficient to run natively on commonplace laptops and smartphones — even in internet browsers.
Lin, a PKU humanities graduate and polyglot, has lengthy advocated for this “algorithm-hardware co-design” to bypass compute constraints—a philosophy he detailed on the January 2026 Tsinghua AI Summit.
For the developer neighborhood, Qwen3.5 wasn’t simply one other replace; it was a blueprint for the “Agentic Inflection,” the place fashions shift from being chatbots to autonomous “all-in-one AI staff” able to navigating UIs and executing advanced code.
The enterprise dilemma
For the 90,000+ enterprises presently deploying Qwen by way of DingTalk or Alibaba Cloud, the management vacuum creates a disaster of confidence.
Many corporations migrated to Qwen as a result of it supplied a “third approach”: the efficiency of a proprietary US mannequin with the transparency of open weights.
Alibaba has lately consolidated its AI efforts into the “Qwen C-end Enterprise Group,” merging its mannequin labs with client {hardware} groups. The objective is obvious: transition Qwen from a analysis mission into the working system for a brand new period of AI-integrated glasses and rings.
Nevertheless, the reported appointment of Hao Zhou, a veteran of Google DeepMind’s Gemini crew, to steer the Qwen crew signifies a shift from “research-first” to “metric-driven” management.
Trade analysts, together with these cited by InfoWorld, warn that as Alibaba pushes to fulfill investor calls for for income development, the “open” in Qwen’s open-weight fashions could develop into a secondary precedence — just like what we noticed with Meta after the disappointing launch of its Llama 4 AI mannequin final spring, and subsequent reorganization of its AI division, seeing the hiring of Scale AI co-founder and CEO Alexandr Wang and following departure of preeminent researcher Yann LeCun.
Enterprises counting on the Apache 2.0-licensed Qwen fashions now face the chance that future flagships —such because the rumored Qwen3.5-Max—can be locked behind paid, proprietary APIs to drive Cloud DAU (Each day Energetic Person) metrics.
The takeaway? When you worth Qwen’s open supply efforts, obtain and protect the fashions now, when you nonetheless can.
The “Gemini-fication” of Qwen?
The interior friction at Alibaba mirrors the tensions seen at OpenAI and Google: the “soul” of the machine is commonly at odds with the “scale” of the enterprise. Xinyu Yang, a researcher at rival Chinese language AI lab DeepSeek, captured this sentiment in a stark put up on X: “Change the superb chief with a non-core individuals from Google Gemini, pushed by DAU metrics. When you choose basis mannequin groups like client apps, don’t be shocked when the innovation curve flattens.”
This “Gemini-fication”—the shift towards a extremely regulated, product-centric tradition—threatens the very agility that allowed Qwen to surpass Meta’s Llama in by-product mannequin creation. For the worldwide AI neighborhood, the lack of Junyang Lin is symbolic.
He was the first bridge between China’s deep engineering expertise and the Western open-source ecosystem. With out his advocacy, there are fears that the mission will retreat right into a “walled backyard” technique just like its Western rivals.
‘Leaving wasn’t your selection’
The technical brilliance of the Qwen3.5 launch has been overshadowed by the heartbreak of its creators. On social media, the sentiment among the many crew members who constructed the mannequin is one in every of mourning moderately than celebration:
Chen Cheng, a Qwen contributor, explicitly alluded to a compelled departure, writing in a put up on X: “I am actually heartbroken. I do know leaving wasn’t your selection… I truthfully cannot think about Qwen with out you.”
Li instructed the exit signaled the tip of broader ambitions, equivalent to a deliberate Singapore-based analysis hub: “Qwen might have had a Singapore base, all because of Junyang. However now that he is gone, there is not any purpose left to remain right here.”
What occurs to Qwen’s open supply AI efforts from right here on out?
The identified information are easy: Qwen has by no means been technically stronger, but its founding core has been dismantled. As Alibaba prepares to face traders for its fiscal Q3 earnings report on March 5, the narrative will seemingly concentrate on “effectivity” and “industrial scale.”
For the enterprises presently excited in regards to the 60% value reductions promised by Qwen3.5, the fast future is vivid.
However for the bigger AI neighborhood, the price of that effectivity would be the lack of probably the most vibrant open-source lab within the East.
As Hao Zhou takes the reins, the world is watching to see if Qwen stays a “mannequin for the world” or turns into merely a element in Alibaba’s company backside line.
