Qwen dominates open models
Alibaba’s Qwen family now accounts for more than half of global open‑source model downloads, a shift that suggests the open ecosystem’s centre of gravity is moving east. That matters because download share is a proxy for mindshare, tooling momentum and where engineers experiment when they need portability or self‑hosting options. (scmp.com)
One company’s models are now being downloaded more than the rest of the open model world put together. A South China Morning Post report, citing tracking firm Interconnects AI, says Alibaba’s Qwen family passed 50 percent of global open-source model downloads as of March 2026 and reached 942.1 million cumulative downloads. (scmp.com) That is a sharp jump from January 2026, when the same Qwen family was reported at 700 million downloads on Hugging Face. In roughly two months, the total grew by more than 240 million, which helps explain how one model line could suddenly tower over Meta’s Llama and DeepSeek. (scmp.com) A model download is not the same thing as a paying customer, but it is where habits form. When engineers pull a model from Hugging Face or ModelScope, they usually also pick its prompts, wrappers, inference stack, and fine-tuning recipes. (huggingface.co, alibabacloud.com) Qwen’s rise did not come from one giant model. Alibaba kept shipping a ladder of sizes, from small versions that fit on a laptop to larger mixture-of-experts systems, so a student, a startup, and a cloud team could all stay inside the same family. (github.com, qwen.ai) Alibaba also made the licensing unusually permissive for a company of its size. The official Qwen3.5 repository on GitHub is under Apache 2.0, which lets developers modify and commercialize the weights without the tighter usage limits that slowed adoption of some rivals. (github.com) The release cadence mattered too. Alibaba announced Qwen3.5 on February 16, 2026, added medium models on February 24, and shipped smaller models on March 2, which meant new checkpoints kept landing fast enough to stay on developers’ front pages. (github.com, alibabagroup.com) The models also widened their addressable market. Alibaba said Qwen3.5 supports 201 languages and dialects, up from 119 in the Qwen3 series, which gives it a better shot in markets where English-only tools feel imported and incomplete. (alibabagroup.com) There is also a distribution trick here that Western open-model companies do not fully match. Qwen is available through Hugging Face, GitHub, ModelScope, Qwen Chat, and Alibaba Cloud Model Studio, so the same model can spread through both global developer channels and Alibaba’s own stack. (github.com, alibabagroup.com, alibabacloud.com) That helps explain why this is bigger than a leaderboard story. If the most-used portable models now come from Alibaba, then the default open tools for coding assistants, internal chatbots, and self-hosted agents may increasingly inherit Chinese model choices even when the final app is built in the United States or Europe. (scmp.com, huggingface.co) Meta’s Llama still matters, and DeepSeek still has breakout mindshare, but download gravity is moving. In open models, the company that gets installed first often becomes the one developers optimize around later, and right now that company is Alibaba. (scmp.com, scmp.com)
