At 2 a.m. on March 4, 2026, Junyang Lin posted eight words on X:
"me stepping down. bye my beloved qwen."
Two days earlier, he had open-sourced Qwen3.5 — a release Elon Musk publicly praised for its "intelligence density."
Five months later, Qwen3.8-27B shipped with DeepSWE jumping from 13.3 to 42.2, a 3x gain in a single generation.
When people trace that result back, one question keeps coming up: why did the man who built Qwen leave when it was doing better than ever?
From Search and Recommendation to Qwen Lead
In 2019, Lin joined Alibaba's DAMO Academy as an algorithm engineer in the intelligent computing lab, working on NLP for search and recommendation.
In late 2022, Alibaba merged DAMO's language and vision teams into Alibaba Cloud and created the Tongyi Lab. Lin, at 29, was put in charge of the Qwen project — the chance to define a domestic open-source model family from zero.
In April 2023, Tongyi Qianwen (Qwen) launched.
A Billion Downloads and the Youngest P10
The numbers after that speak for themselves.
Per Caixin, Qwen models passed 1 billion downloads with over 200,000 derivative models — one of the deepest ecosystems in open-source AI.
In 2025, Lin was promoted to P10 at age 32 — the youngest in Alibaba Cloud's history.
His final timeline reads strangely: in January 2026 he was speaking at industry forums about reinforcement learning. In late February he was still posting recruiting messages. On March 2, Qwen3.5 shipped. On March 3, his resignation was processed. On March 4, the X post.
A man recruiting for his team two days before leaving doesn't walk away on impulse. He walks away at the end of a long disagreement.
The Split: Integrated, or Separated?
Caixin traced the departure to a set of strategy disputes, the sharpest being: should pretraining and post-training be integrated, or separated?
Lin favored integration — one team owning both pretraining and post-training, base capability and alignment capability wired together.
The market trend favored separation: ByteDance's Doubao team has been splitting, and "pretraining teams" versus "post-training teams" are becoming two different species across the industry.
This is not office politics. This is the exact question every major model lab was answering in 2026: where does raw capability come from, where does alignment come from, and who hands what to whom?
Lin's second ask got less attention: he wanted the Qwen team to control its own AI infrastructure. Alibaba Cloud, meanwhile, was pushing productization — models into vertical industries, research serving commercial deployment.
Jingren Zhou's line was "the model and the cloud can't be separated." Lin's answer was to leave.
Neither side was obviously wrong. But the disagreement was real enough to move a P10.
After the Exit: A Two-Role Structure
Alibaba's answer was to split one person's job into two.
Strategy went to Jingren Zhou — USTC gifted-youth class, Columbia PhD, eleven years at Microsoft, Alibaba Cloud chief scientist, CTO, and as of December 2025 an Alibaba partner. Since March 2026 he has overseen the Qwen program. His thesis fits in one sentence: the model is the product entry point of the cloud.
Technical leadership went to Hao Zhou — a former senior DeepMind researcher who took over as Qwen tech lead in March 2026. His public résumé is short. His output is not: Qwen3.6 and Qwen3.8 both shipped on his watch.
Organizational theory has a plain name for this: de-risking the bus factor.
The Inheritance Question
Here is the detail most people miss: Qwen3.8-27B's architecture is not a fresh start by the new team. It is a continuation of Qwen3.5.
The 3:1 hybrid linear attention, the natively multimodal foundation, the 262K context — all of it traces back to Qwen3.5, the last release Lin open-sourced before leaving.
What the new team did was amplify: DeepSWE 13.3 → 42.2, OSWorld 63.9 → 84.3, SWE-bench Pro 61.7 — first place on the table.
So the precise answer to "who built Qwen3.8" has three layers: the architecture is Lin's legacy, the engineering amplification belongs to Hao Zhou's team, and the strategic and commercial frame belongs to Jingren Zhou.
No model's final form is one person's work. But when it delivers a step change, people ask where the founder went.
Coda
What makes "bye my beloved qwen" hold up is that it doesn't read like a resignation statement.
No thanks to the company. No pitch for the future. Just an engineer saying goodbye to something he built for seven years.
The message above it was a WeChat Moments post titled "A Toast to Myself." The message below it was Alibaba's public commitment, on March 5, to keep supporting large models and the open-source strategy — followed by Qwen3.6 and Qwen3.8 arriving on schedule.
China's first high-profile open-source founder departure ended like this: the person left, the architecture stayed, and the cadence never broke.
For open-source ecosystems, that may matter more than any single technical result. A mature open-source system has to survive the exit of its founder. Qwen spent five months proving it could.
And the P10 engineer left the industry with more than a billion-download model. He left the question every AI organization eventually has to answer:
When the split comes, do you integrate — or separate?
