Why it matters
Qwen is the model family Alibaba uses to compete with the frontier labs already on this site, and it is the one most often downloaded when people want an open-weight alternative. It is not a startup. The August 3, 2026 press note describes Qwen as the AI technology behind Alibaba's own services, and the public site is qwen.ai.
The current flagship, Qwen3.8-Max, is a 2.4 trillion parameter mixture-of-experts model with 95 billion parameters active and a context window of up to 1 million tokens. Alibaba put it on Model Studio on August 3 and released the weights shortly after. A 27 billion parameter sibling followed under Apache 2.0.
A model family, not a company
Wikipedia dates the public beta to April 2023, under the name Tongyi Qianwen, with general availability that September after regulatory clearance. Development is credited to Alibaba Cloud. There is no Qwen listing and no Qwen revenue line in the documents opened here.
Leadership moved in 2026. Yicai reported on September 21 that Liu Dayiheng had been named to lead the Qwen LLM team, citing Alibaba Cloud's Apsara Conference site, and that former head Lin Junyang had left in March. The next day's roadmap was delivered by group chairman Joe Tsai and group CEO Eddie Wu. Liu is the project lead in that report. He is not the CEO of Alibaba.
What shipped in August 2026
Qwen3.8-Max was the API flagship. The August 17 post then said the flagship weights, Qwen3.8-2.4T-A95B, were on Hugging Face and ModelScope, and that this was the first time a Qwen-Max-class model had been released that way. Qwen3.8-27B is the small multimodal model Alibaba said can run on a laptop after quantization.
The licenses are not the same. Alibaba said the 27B model is Apache 2.0. Wikipedia says the 2.4 trillion parameter release requires a commercial license once a provider generates more than US$50 million in 12 months. Treating every Qwen weight as unrestricted open source would be wrong.
Downloads are not revenue
Alibaba's August 17 post claims more than 460 open models, over 300,000 derivatives, and more than 3 billion global downloads. In the same post it quotes a Hugging Face summer 2026 note for a different pair of numbers: 151,448 Qwen-based repositories, and 2,045 million downloads from January to July 2026 among repos with a declared parameter count. Those are not one statistic.
Neither number is sales. No Qwen profit, price list, or Alibaba Cloud segment revenue was opened for this profile. The business channels that were named are the Model Studio API and the published weights.
What to watch
On September 22, 2026 Alibaba said Qwen 4 is in training, and that Qwen 4.5 and Qwen 5 are projected to reach 5 to 10 trillion parameters. That is a company projection, not a release. Eddie Wu also set a 2032 target for Alibaba Cloud data-center capacity above 20 gigawatts. That target belongs to the cloud group, not to a Qwen income statement.
The other thing to watch is who actually runs the model line after Lin Junyang's March departure. Yicai's account depends on the Apsara Conference site. If Alibaba publishes a different title, the project-lead line should move with it.