Alibaba released Qwen3.8-Max on Monday, calling it the most capable model it has ever built. The weights land on Hugging Face and ModelScope next week—the first time the company has given away a model at Max scale.
The specs are big: 2.4 trillion parameters total, with 95 billion switched on at any moment. Parameters are the number of dials a model is able to handle.
This is extremely important for efficiency as it means it would not require too much resources to run. Think of it as a huge library where only the relevant shelf lights up for each question. Small businesses and labs with good enough hardware are now able to run a state-of-the-art model without spending the same as a huge datacenter.
Built to run inside a rival's toolsQwen3.8-Max ships with instructions for Claude Code and Codex, the coding tools made by Anthropic and OpenAI. Alibaba's API speaks both companies' protocols. Most of its coding benchmarks were run inside Claude Code.
And those benchmarks don't flatter it. Across 31 text tests, Anthropic's Fable 5 takes 15 first place spots, OpenAI's GPT-5.6 Sol takes nine, Qwen takes seven. On the 12 coding tests, Qwen wins exactly one. However, in terms of intelligence costs, this model is extremely cheap and efficient, which means that even if it requires more iterations or reasoning, the cost of getting the job done will be much lower, nearly 30% of what Claude Fable 5 charges.
That said, flip to multimodal work—documents, video, spatial reasoning—and the rankings invert. Qwen leads most of that table.
That door is now open, and the timing isn't an accident. Chinese open-weight models went from under 2% of tokens on OpenRouter in late 2024 to roughly 61% by mid-2026. Qwen passed Meta's Llama as the most self-hosted model in the world.
So Alibaba is losing on paper and winning on distribution. If you can download something that comes close for free, second place is a fine place to be.
