Alibaba Opens the Weights of Qwen3.8-Max, a 2.4 Trillion Parameter Model

Alibaba published the weights for Qwen3.8-Max on Hugging Face on August 13. It is the first Max-tier model in the family to go open-weights.

The headline number is large: 2.4 trillion parameters.

That number isn't the whole story

The architecture of this LLM is Mixture-of-Experts. So while the full model holds 2.4 trillion parameters, only around 95 billion activate for any given query.

That distinction matters. The model delivers the quality of something enormous while the compute cost per query stays far below what a dense 2.4T model would demand.

The context window is 1 million tokens, and the model accepts text, image and video input.

The problem: who can actually run it?

This is where enthusiasm meets reality. "Open weights" means you can download it, not that you can serve it.

Standing up a model at this scale requires a cluster of expensive GPUs. The number of teams worldwide who can realistically do that is small.

Which is why Alibaba also shipped a smaller checkpoint: Qwen3.8-27B. That one runs on ordinary hardware and is probably what most people will actually use.

The license changed

A detail that is easy to miss: this model did not ship under Apache 2.0.

Both Qwen3.8 and Kimi K3 have moved to custom licenses that impose additional authorization thresholds on hyperscale commercial products and MaaS businesses.

For research and small projects, nothing changes. If you intend to build a large commercial service on top of it, read the license text.

The bigger picture

What is striking is how quickly Chinese labs have converged on a shared template: a huge MoE model, a million-token context, open weights, and a license that keeps large-scale commercialization on a leash. Zhipu's GLM-5.3 and DeepSeek's new harness follow the same path.

That is precisely the point where US labs keep their weights closed. 🤖