Reddit上关于Qwen人事变动的讨论

看上去是因为高管对Qwen团队现有的结果和生态位都不满意导致的。
原文链接:
https://www.reddit.com/r/Qwen_AI/comments/1rkmdry/junyang_lin_leaves_qwen_takeaways_from_todays/

The original Qwen team of over 500 people was constantly demanding more funding and more GPUs, yet they operated without any KPI evaluations.

Ultimately, their results were inferior to the small models cleverly distilled by MiniMax, despite Qwen’s total burn rate (costs) being more than 10x higher.

To the executives, the whole operation was a “black box” they couldn’t influence. Their only role was to provide whatever funding, headcount, or hardware was requested.

Looking at the final DAU (Daily Active User) metrics, the executives could only watch in helpless frustration.

At that point, the boss brought in someone from DeepMind as an observer. Their conclusion was equally damning: “The output looks like a temporary toy made by an intern”—hardly a glowing review.

In response, the boss began breaking down metrics into sub-indicators to prevent “self-congratulatory” reporting.

The team leaders interpreted this move—breaking down metrics and setting KPIs—as a threat to their positions. They attempted to leverage a collective resignation as a threat.

And so, it played out: “If you want to quit, then quit…”
Meeting takeaways:

⁠HR’s Spin: The Chief HR Officer is framing these changes as a way to bring in more talent and resources, not as a downsizing or a setback.

⁠The "Big Picture": Management says Alibaba is now a "model company." Qwen isn't just a side project for the base model team anymore—it’s a Group-wide mission. They want a "closed-loop" system to move faster, but they admitted they communicated the new structure poorly.

⁠The "Price" of Growth: Because Qwen is the top priority, the team has to expand, which means the "formation" has to change. They basically said, "Growth isn't free—there’s always a price to pay."

• The Leadership Drama: They argued that while relying solely on Junyang’s brain is efficient, Jingren had to figure out where to put Zhou Hao to make things work. They claim there was no “office politics” involved. (Interestingly, management previously claimed Zhou Hao asked to report to Jingren because he was worried about fitting in).

  1. Scaling Pains: They argued that 100 people aren’t enough for a project this big. They need to scale up, and in that process, they “can’t please everyone.”

  2. Eddie Wu’s Defense: Eddie (Wu Ma) blamed the resource shortage on China’s unique market conditions. He apologized for not being aware of the resource issues sooner, but insisted he’s the most aggressive CEO in China when it comes to hunting for computing power. He claims Qwen is his #1 priority.

  3. The “Bottleneck” Excuse: When asked why the Group was “strangling” their resources, Eddie claimed he had no idea there was a block. He said the priority was always high and blamed the whole thing on a “breakdown in communication.”

  4. Jingren’s Take: Jingren admitted resources have always been tight. He even claimed that he’s being “sidelined” or bypassed himself. He also acknowledged the long-standing internal complaint that Alibaba Cloud’s own infrastructure is a pain to use, calling it a “historical issue.”

  5. The Final Word on Junyang: When someone asked if Junyang could come back, the HR Lead shut it down. They said the company won’t “put anyone on a pedestal” or pay “any price” to keep someone based on “irrational demands.” They then turned it on the audience, asking, “What do you all think your price is?”

The Bottom Line: Management is prioritizing the “Group” over individual stars. They are essentially telling the team that if they want to be part of the “big mission,” they have to accept the new hierarchy and the loss of key leaders.

11 个赞



10 个赞

看起来情况比想像的糟糕?
别的团队花同样的资金可以搞出 200B, 300B 甚至 700B 的 SOTA LLM
Qwen 却只能死守 SLM
然后还有内部斗争 :distorted_face:

9 个赞

看通义大会的新闻报道就知道了

哪有开会 HR 上去说话的

16 个赞
4 个赞

确实比较弱,Qwen 生图 2.0 Pro 已经上架 fal 了,没有任何波澜,而看见那示例的 Demo 图我连为它写个调用脚本都懒得。

4 个赞

有一定的道理,qwen拿着最多的钱和资源到现在却没有什么标杆模型,现在这qwen3.5也只能勉强算一个吧。

别的模型一眼就能看出哪些是主线大模型哪些是小模型而且也只专心维护升级那几个模型,成本自然低,而qwen混乱的产品线不去研究根本不知道他到底哪些模型是干什么的,整这么多模型成本炸裂不说大部分都没人在乎

我觉得以后qwen可能是整一个顶级闭源大模型赚钱,一个中等开源模型以及小模型抢开源市场

7 个赞

一堆模型我觉得不影响什么
问题是 Qwen Max 为啥那么平庸

3 个赞

刚好看到这篇。Reddit上都一片哀嚎,觉得开源社区再也没法赶上闭源
没必要悲观,总是有新的模型出来的

主要还是小模型,其他的开源模型肯定会有的,glm kimi minimax 也都开源。但是一般人想在本地跑就很难了

你不说我都不知道fal上架了, 估计不开源了.

qwen3.5小模型领先同行至少一年, 以前google/微软还发一下开源小模型, 现在直接退出群聊了.

1 个赞

我也要哀嚎了:sob:不过不是因为qwen
reddit给shadowban完了和似了差不多
发小作文快半个月一点动静没有:anxious_face_with_sweat:

1 个赞

你做了什么?reddit对ip要求还蛮高的

那我感觉是ipban了,580左右kam都能给我干封了。
可能连发两帖导致的,应该判断发太快了ip不干净

唉 我们也不是内部人也不知道实际情况
今天还看到流出另一种说法的

但是glm minimax那波的确是 虽然qwen3去年年中的确很出彩
但是在竞对乱杀 迭代了好几次的时候 qwen去哪了:joy:
(我这个说法有点马后炮了

Qwen 的产品线确实很乱,阿里云百炼上各种模型眼花缭乱,想用根本不知道怎么选,尤其是即使同代模型,不仅有开源的几款,还有不开源的几款,而且价格也不低

2 个赞

感觉他们内部很乱啊

大公司当然不必小公司,但是同样的土地,你就是比别人差咋办嘛,和字节比大厂比日活,和小厂比性能,你说你是团队负责人结果说话,你咋说呢。哈哈哈

1 个赞

这算不算大公司病呢,会不会重复meta的路子啊……快进到乱收购、乱换人 :sweat_smile: