AI Pulse · AI Trends Pulse

Alibaba just dropped Qwen3.8-Max, a 2.4 trillion parameter model that uses mixture-of-experts architecture to keep about 95 billion parameters active at any given time. It handles a one million token context window. The bigger news is that Alibaba plans to open-source the weights next week, the first time the company has released a model at this scale for anyone to download and run. Markets noticed. Alibaba’s stock jumped 4.5% in premarket trading in New York and 7% on the Hong Kong exchange, according to the original report.

Why this matters for your business

Open-source weights mean you can run this model on your own infrastructure if you have the capacity, or through a provider that hosts it. You’re not locked into Alibaba’s API or pricing. For companies that need to process long documents, legal contracts, or large codebases, the one million token context is the practical detail. That’s roughly 750,000 words in a single prompt, enough to feed in multiple reports or an entire project brief without chunking it up.

The mixture-of-experts design keeps compute costs lower than a dense model of the same headline size. Only a fraction of the parameters activate for each task, which matters if you’re running inference at scale. This is the kind of capability that gets built into systems like our Omni Command Centre, where you need flexible model options depending on the task, not just one vendor’s API.

Alibaba is clearly positioning this as a competitor to OpenAI and Anthropic’s frontier models. Whether it performs at that level in practice, we’ll know more once people start testing the open weights next week. For now, it’s another serious option in the mix, especially if you want control over where your data lives.