Alibaba built one of the world’s most widely adopted AI model families by giving it away for free. Now it has to figure out how to make money from it.

The Chinese tech conglomerate launched its Qwen family of AI models in 2023 through its Tongyi Lab, releasing them as open-source tools that anyone could download, modify, and deploy. By the end of 2025, Qwen models had climbed from roughly 1.2% of global self-hosted AI usage to around 30%, surpassing Meta’s Llama in self-hosted deployments.

From open-source to open for business

In 2026, Alibaba made a decisive pivot. The company transitioned its flagship Qwen models, including Qwen3.6-Max and Qwen3.7-Plus, from open-source to a closed, API-only model with usage fees attached. Free tier access for several tools was quietly removed. The playbook is borrowed directly from OpenAI and Anthropic: build an audience with accessibility, then monetize through enterprise API contracts.

Alibaba also released Qwen3.5 on February 16, 2026, a version the company says introduces agentic capabilities, meaning the model can plan and execute multi-step tasks rather than just respond to single prompts. It is also reportedly 60% cheaper to run than earlier versions.

The money behind the strategy

Alibaba committed a ¥3 billion promotional budget for Qwen tied to the 2026 Lunar New Year period, with the AI and cloud infrastructure buildout running through 2027.

The company has been showcasing Qwen prominently across Shanghai trade events in 2026, including appearances on March 12 and at MWC Shanghai in June.

Leadership changes are also in motion as part of this commercialization push, though the company has been characteristically understated about the specifics.

What this means for markets and the AI landscape

Meta continues to release Llama models openly with no direct monetization path, treating it as infrastructure for its ad business. Mistral has taken a hybrid approach, open-sourcing smaller models while keeping enterprise versions proprietary. Alibaba is now firmly in the proprietary camp for its most capable models, which puts it in direct competition with OpenAI and Anthropic for enterprise API contracts.

The risk for Alibaba is that developers who self-host have little incentive to pay for API access when the model weights are already available. Alibaba’s bet is that the newest, most capable Qwen versions, starting with 3.6-Max and 3.7-Plus, will be good enough that enterprises prefer the managed API over the operational complexity of self-hosting at scale.