Alibaba Opens 2.4-Trillion-Parameter Qwen3.8 Weights
Alibaba has released its largest Qwen model’s weights, giving independent operators a frontier-scale foundation for coding and agents.
The release
Alibaba’s Qwen team has published the weights and model card for Qwen3.8-2.4T-A95B, making the foundation on which the hosted Qwen3.8-Max service is based available for outside organizations to download and operate. The mixture-of-experts model contains 2.4 trillion parameters but activates about 95 billion for each token, an approach designed to increase total capacity without paying the computational cost of using every parameter on every request.
The public model has a native 262,144-token context window and is positioned primarily for coding, complex tool use and long-running agent workflows. Unsloth published an extreme one-bit quantization that reduces the roughly 4.9-terabyte full model to about 397 gigabytes. Even that compressed edition remains beyond ordinary consumer hardware, underlining the difference between weights being downloadable and a model being broadly practical to self-host.
Some inference providers separately announced day-zero availability for Qwen3.8-branded services. Those announcements must be evaluated model by model: availability of the proprietary Qwen3.8-Max service does not by itself confirm hosted inference for the downloadable Qwen3.8-2.4T-A95B checkpoint.
The downloadable release also should not be treated as identical to Alibaba’s commercial Qwen3.8-Max product. According to the model documentation, the hosted edition is based on the open model but adds vision input, non-thinking operation, built-in tools and a default one-million-token context. Users evaluating sovereignty or customization therefore receive the core text model, but not necessarily every capability of Alibaba’s managed service.
Why it matters
This is one of the largest publicly downloadable language models yet released. More importantly, it gives cloud companies, national AI programs and large enterprises a new alternative to closed frontier APIs for agentic software work. The hardware demands will limit direct adoption, but the weights create a common foundation for quantization, fine-tuning and systems research—and put additional pressure on Western labs to justify closed access with capabilities that cannot be reproduced around an open model.