Alibaba has just announced Qwen 3.8-Max, its largest language model to date. According to the company, the model has 2.4 trillion parameters, can handle a context of up to one million tokens, and is both multimodal and agent-focused.
What's interesting for us in Europe isn't just the size. It's the promise that the weights will be released as open weights as early as next week.
What Do We Know — and What Don't We?
Alibaba has made the model available via its own APIs and the QwenWork platform, but concrete benchmarks have not yet been published. That means performance claims should be taken with the caveat that comes with any product launch lacking independent verification.
For companies considering using the model in production, two things are worth watching closely: the license terms for the upcoming open weights, and the actual benchmark results from independent parties.
Open Weights Are Changing the Rules for Hosting
Once the weights land as a download, a different discussion opens up than with cloud-only models: Can we run it ourselves? How much hardware does it need? And, not least: Where should it run?
A 2.4-trillion-parameter model isn't something you throw onto a single GPU. But even if only the smallest quantized versions become practically usable for ordinary companies, the launch points in the same direction as Kimi K3 and OpenAI's upcoming open-weight model: the strongest models no longer have to live with American or Chinese cloud providers.
"Open weights let companies choose hosting based on their own requirements — not the vendor's home country."
What Does This Mean for Liviate Customers?
- Data sovereignty: An open-weight model can be run in the EU on your own infrastructure.
- Predictable costs: No per-token billing to external APIs.
- Compliance: Easier to document data flow and processing location.