Alibaba has open-sourced its entire Qwen3 model series under the Apache 2.0 license this year, releasing eight sizes ranging from 0.6B to 235B parameters, all fully available.
Why the Apache license matters
Apache 2.0 is one of the most permissive open-source licenses, meaning enterprises can use the models free of charge for commercial purposes without needing additional licensing agreements. By contrast, Meta's Llama series, while also claiming to be "open source," is effectively "open-weight" with various usage restrictions. Alibaba's move represents the most thorough open-source effort among domestic large-model vendors so far.
Technical highlights
The most interesting design in Qwen3 is its hybrid reasoning mode — the same model incorporates both a thinking mode and a non-thinking mode:
- Thinking mode: deep reasoning, suitable for mathematics, coding, and logical analysis
- Non-thinking mode: fast responses, suitable for everyday conversation
One model handles both scenarios, eliminating the need to switch between a "reasoning model" and a "conversation model," which directly optimizes deployment costs.
Benchmark performance
Qwen3-235B (MoE, with 22B activated parameters) trades blows with DeepSeek R1 and GPT-4o on multiple benchmarks. On the cost side, thanks to the MoE architecture, actual inference overhead is equivalent to that of a 22B dense model.
The smaller versions also deliver surprises — Qwen3-4B matches the previous-generation Qwen2.5-72B on many tasks. A parameter gap of more than 40 times has been closed, an impressive leap in progress.
Alibaba's strategy is clear: use the most aggressive open-source approach to capture the ecosystem. Developers can experiment with Qwen3, build products, and commercialize without licensing barriers. Once the ecosystem is running, the company can monetize through cloud services. This mirrors the playbook Meta used with Llama to build an open-source ecosystem.
Sources: Alibaba Cloud official release, CocoLoop, Qwen GitHub repository