The largest model ever produced by Alibaba is based on a mix of experts and only activates 95 billion parameters per token generated.
Alibaba launched Qwen3.8-Max on August 3, 2026, after a preview presentation at the Shanghai World AI Conference in July. The model totals 2,400 billion parameters but only activates 95 billion per token, thanks to a parsimonic architecture with mixed experts that limits the cost of inference despite the announced size.
Multimodal, Qwen3.8-Max accepts text, images and video input, with a context window of one million chips and up to 131,072 chips produced per response. The API access rate is set at $2 per million input chips and $6 per million output, which is significantly more aggressive than the US border models.
The most commented element remains the announced publication of weights on Hugging Face and ModelScope, accompanied by a lighter model, Qwen3.8-27B: this would be the first time that Alibaba has opened the weights of a model in its Max range. The reference results provided by the company place it in front of several Western competitors on the OSWorld-Verified test, but these measures are self-reported and independent verification remains limited within days of release.
Read the full article on the original source
Go to the source →