Alibaba released a new artificial intelligence model called Qwen3.8 Max on 3 August and priced it well below what Anthropic and OpenAI charge for their best systems. The company also promised something unusual for a model this large: to publish it free, so anyone can download it and run it on their own machines. Alibaba shares closed around 7 per cent higher in Hong Kong that day, with some outlets reporting a rise closer to 6 per cent.
That free release has not actually happened yet, despite a wave of headlines saying it has. And for Australian businesses there is a second catch nobody has written about: there is no Alibaba data centre in this country to run the model from.
What Alibaba actually released
Qwen3.8 Max is a very large system, built from 2.4 trillion internal settings, though it switches on only a small fraction of them for any given request. That design is what keeps a model this size affordable to run. It takes text, images and video, and holds about a million tokens at once. Tokens are the chunks of text a model reads, so that is a great deal of material in one go.
You can use it today through Alibaba Cloud’s own service, through the company’s new work assistant and through OpenRouter, as reported by the South China Morning Post. What you cannot do is download it. Alibaba says it will publish the model free next week, the first time it would have done that with its largest system. As of 4 August nothing had appeared on Hugging Face, the site where those releases are posted.
The scores do not agree
Alibaba published a table of test results putting its model at or near the top. On one measure of how well a model handles long technical tasks, it scored 86.6 against 84.6 for two of Anthropic’s Claude models, although OpenAI’s GPT 5.6 Sol beat all three at 88.8. On the main programming test in the same table, Alibaba’s model scored 67.7 against 80.0 for Claude Fable 5.
The footnotes matter. Alibaba tested its own model with a setup it built itself, while taking rival scores from published leaderboards, and it ran the programming test on a version it had corrected. About a third of the table is made up of tests Alibaba invented, which nobody outside the company can repeat. Its own note underneath warns that the Claude Fable 5 results may involve fallbacks.
What is missing is an independent scoreboard. As of 4 August the benchmarking site Artificial Analysis had not published a page for this model, so every figure in circulation traces back to Alibaba. The one concession in its own material is on Frontend Code Arena, where it puts Qwen3.8 Max at 1,668, some 37 points behind the strongest configuration of Claude Opus 5. Treat the rest as vendor marketing until somebody else runs the tests.
What it costs, and what it really costs
The price is the headline. Alibaba charges $2 per million tokens going in and $6 per million coming out, listed at the same rates on OpenRouter. It also built the service to answer the same commands as Anthropic’s, so a development team already using Claude can point its tools at Alibaba instead by changing three settings. The barrier to a genuine trial is close to nothing.
The rate card is not the bill. A million token context and $6 per million on the way out add up quickly once an agent is running long jobs unattended, and nobody outside Alibaba has published throughput figures yet. The same trap showed up when Kimi K3’s free release landed in July. Price a real week of work, not the rate card.

Why Australian businesses should check the map
Alibaba runs this service from six places: Beijing, Singapore, Frankfurt, Tokyo, Hong Kong and Virginia in the United States. None is Australian. Alibaba Cloud shut its Australian data centres on 30 September 2024, eight years after opening in Sydney, and told customers to move to Singapore or elsewhere in Asia. A business here that feeds customer records into this model is sending them overseas, because there is no local option to pick.
That matters under the Privacy Act. Australian Privacy Principle 8 puts the responsibility on the organisation handing the information over when it goes to an overseas recipient, and that organisation generally stays accountable for what happens next. If you sell to government, or hold health or financial records, this is a conversation for before the trial rather than after it. Our guide to AI data residency goes into the detail.
There is form for how Canberra responds. The Department of Home Affairs told federal agencies to remove DeepSeek, another Chinese AI product, from government systems on 4 February 2025, and the Group of Eight universities, NBN Co, the ABC, Australia Post and Optus followed. That order covered the hosted service, not software running on your own machines, which is exactly why next week’s free release changes the question rather than repeating it.
What to watch next week
The free release is the part worth waiting for, and not only because independent labs could finally check Alibaba’s numbers. Software you can run on your own machines is the only way to keep a model of this class inside an Australian data centre, and the licence decides whether you are allowed to. Alibaba has published neither a date nor the terms.
Running a 2.4 trillion setting model yourself is beyond almost every medium sized Australian firm anyway. A smaller companion version, Qwen3.8 27B, is not, and that is the one most operations teams should watch for. The sensible move this week is a costed trial on data that does not matter, not a switch.
Cheap capability at this level is worth taking seriously. Whether it is worth taking on depends less on the test table than on where your data ends up and who answers for it when it does, which is the part of AI strategy work that tends to arrive last.
