The Story of Qwen: Alibaba’s AI Models From 7B to 2.4T

MarkTechPost’s Asif Razzaq traces Alibaba’s Qwen from an invite-only enterprise chatbot in April 2023 to the August 2026 open release of Qwen3.8-2.4T-A95B, with 2.4 trillion parameters and 95 billion active. Along the way: Qwen2’s move to Apache 2.0, Qwen3’s hybrid thinking across 119 languages, the QwQ-32B reasoning model that claimed parity with a far larger DeepSeek-R1, and four generations in roughly six months of 2026.

The scale of adoption is the headline. By March 2026, Reuters put Qwen at more than 400 open models and over a billion downloads, and Alibaba said Qwen was the first open family to pass 200,000 derivative models on Hugging Face. AI Singapore switched its Sea-Lion base from Llama to Qwen, and CNBC reported Qwen would be integrated into Apple Intelligence in China.

Qwen is now default infrastructure for a large share of the world’s open-model builders. How Alibaba licenses it shapes what everyone else can build.

We think Qwen is the clearest case study of open weights as strategy. Alibaba shipped fast, shipped everything from tiny to huge, and moved most of it to Apache 2.0 when that was the fastest way to win developers. It worked. Qwen didn’t just compete with Llama — in a lot of places, it replaced it. For builders, that has been a genuinely good deal: strong, permissive models that run on hardware real teams can afford.

But the license table in this history tells the more important story. Flagships went API-only starting with Qwen3-Max in 2025. In 2026 the Plus and Max tiers stayed proprietary. The 2.4T open release came with a catch: providers earning over US$50 million in 12 months need a commercial license. And the latest image model is research-only. The pattern is consistent — permissive at the small and mid-size end to keep developers building, closed or conditioned at the frontier to earn money.

That’s a rational business. It’s also a warning for anyone who has built a product on the assumption that open will stay open. The terms can shift with each generation, and a company’s needs change when the revenue question gets sharper. The leadership exits in March — tech lead Junyang Lin and other senior figures, amid a reorganization into horizontal teams — add uncertainty about how the culture that produced this cadence holds up.

I’m watching Qwen 4. Alibaba says it is in training, with Qwen 4.5 and Qwen 5 projected at 5 to 10 trillion parameters, but no date, sizes or license. The license will tell us more than the parameter count.

The two-tier split mirrors a broader shift: the South China Morning Post reported Chinese AI giants pivoting toward proprietary models for revenue. Alibaba CEO Eddie Wu said in 2025 that AI and cloud spending would exceed a RMB 380 billion three-year plan, and in March 2026 Alibaba consolidated its AI units under a group he leads.