Alibaba just released Qwen3.8, its newest and largest AI model, and for the first time made a version of its top-tier model available to download for free. That top-tier line, called Max, has always been Alibaba's premium, closed product. Giving it away changes the calculation for any business shopping around for AI tools.
The full model has 2.4 trillion internal settings, but only 95 billion switch on for any single task, like a huge staff of specialists where only a small team gets pulled in per job. That design keeps the cost of running it down, even though the model itself is roughly six times larger than Alibaba's previous flagship.
On tests covering coding, research, and long multi-step work, Qwen3.8-Max lands close to Claude Opus and OpenAI's latest flagship model, sometimes ahead, sometimes behind, depending on the task. It is not the best model on every measure, but it is now firmly in the same tier as the most expensive options from OpenAI and Anthropic.
Alibaba's hosted version charges 2 dollars per million tokens of input and 6 dollars for output, while OpenAI's comparable flagship charges 5 dollars and 30 dollars. That means Alibaba is undercutting OpenAI by roughly 60 percent on input costs and 80 percent on output costs, for a model that performs in the same range.
This is part of a pattern. Chinese AI labs, including DeepSeek and Moonshot AI, have repeatedly released capable models cheaply or for free to win developer attention worldwide, forcing Western labs to defend prices that once looked untouchable. Alibaba is now doing the same thing at its largest scale yet.
Two catches matter here. The free downloadable version is text only, cannot read images, and its working memory is a quarter of what the paid hosted version offers by default. Developers on Hugging Face publicly complained that Alibaba's marketing promised more than the free version actually delivers.
Second, Alibaba added a new rule to the license. Any company earning more than 50 million dollars over a 12 month period from a product built on this model must sign a separate paid agreement and share revenue with Alibaba. Small businesses, individual developers, and researchers are not affected.
There is also a practical limit most readers will not hit directly: running this model yourself requires racks of the newest, most expensive AI computer hardware. Almost nobody will self-host it. Instead, it will show up as a cheaper option inside the cloud platforms and software tools businesses already use.
The takeaway is simple. Expect the AI tools in your stack to get cheaper over the next few months as this kind of competition plays out. If your vendor switches to a model like this one, ask which version they are running and whether your business has crossed the revenue line that triggers a licensing fee.