Alibaba Releases Qwen3.8-Max Weights on Hugging Face

Alibaba has released the weights for Qwen3.8-2.4T-A95B, the open-weight model underlying its flagship Qwen3.8-Max, on Hugging Face. The release marks the first time Alibaba has made weights from a Max-class Qwen model publicly available.

The model became available on August 12, 2026, following Alibaba’s commercial launch of Qwen3.8-Max earlier in August. Developers can access the weights through Hugging Face and deploy the model using established inference frameworks such as vLLM and SGLang.

The release is significant because Alibaba had previously kept its highest-end Max-class model proprietary, while continuing to offer other Qwen models through more open approaches.

A 2.4-Trillion-Parameter Mixture-of-Experts Model

Qwen3.8-Max is built using a mixture-of-experts architecture containing approximately 2.4 trillion total parameters, with around 95 billion parameters activated per token. This approach allows the model to contain a very large parameter pool without activating the entire system for every token processed.

The released Qwen3.8-2.4T-A95B model has a native context length of 262,144 tokens, which Alibaba says can be extended to slightly more than one million tokens in appropriate configurations.

However, it is important to distinguish the downloadable weights from the commercial Qwen3.8-Max service. Alibaba’s own Hugging Face model card states that the commercial Max version adds features including vision input, non-thinking support, a one-million-token context length by default and built-in tools.

Therefore, the open-weight release should not be described as an identical copy of every capability offered by Alibaba’s hosted Max service.

Strong Performance Across AI Tasks

Qwen3.8-Max has attracted attention for its performance on several evaluations.

The model has demonstrated strong results in areas including coding, research, office work and long-running agent tasks. On the OSWorld-Verified benchmark for autonomous computer-use tasks, reports put Qwen3.8-Max at 86.1, compared with 85.0 for Anthropic’s Claude Fable 5 in the cited comparison.

These results indicate that Alibaba’s model is competitive with advanced AI systems on particular evaluations. They should not, however, be interpreted as proof that Qwen3.8-Max is universally superior to every competing model.

Performance can vary significantly depending on the benchmark, model configuration and task being evaluated.

Developers Can Use Standard Inference Tools

One of the most important aspects of the release is accessibility for developers.

The Qwen3.8-2.4T-A95B weights are compatible with inference platforms including vLLM and SGLang, making it possible for organizations with sufficient computing infrastructure to experiment with self-hosted deployment.

Nvidia has also published information on serving the model using its GB300 NVL72 platform, while AMD has announced Day-0 support for the Qwen 3.8 family on several Instinct GPU platforms.

Despite these deployment options, the model’s enormous size means that running it locally is still substantially more demanding than deploying smaller AI models.

The Open-Weight Release Has Important Differences

The release also comes with limitations.

The downloadable Qwen3.8-2.4T-A95B is primarily a text model, unlike the commercial Qwen3.8-Max service, which provides multimodal capabilities. The native context window is also 262,144 tokens rather than the one-million-token default offered by the hosted version.

The model operates with thinking enabled by default, and the open-weight version does not provide all of the configuration options available in the commercial service.

These differences make it important for developers to examine the model card and technical documentation before assuming that the open release provides the same experience as Alibaba Cloud’s API.

Licensing Matters for Commercial Users

The availability of weights does not necessarily mean unrestricted commercial use.

The Qwen3.8-Max weights are distributed under Alibaba’s specified licensing terms, which developers and businesses need to review before deployment. Reports have also highlighted potential revenue-sharing requirements for certain large commercial applications.

Consequently, companies considering commercial deployment should evaluate the applicable licence rather than treating the release as equivalent to an unrestricted permissive licence.

A New Step in China’s AI Competition

Alibaba’s decision comes as competition among Chinese AI models intensifies.

Moonshot AI’s Kimi K3, DeepSeek’s latest models and other large systems have increased pressure on both Chinese and international AI developers.

By releasing weights for a Max-class model, Alibaba is giving researchers and developers greater access to one of its largest AI systems while maintaining a separate commercial version with additional capabilities.

A smaller Qwen3.8-27B model is also expected to expand access to the Qwen 3.8 family for developers with more limited hardware resources.

Frequently Asked Questions

What is Qwen3.8-Max?

Qwen3.8-Max is Alibaba’s flagship AI model, based on a 2.4-trillion-parameter mixture-of-experts architecture.

What is Qwen3.8-2.4T-A95B?

It is the open-weight model released by Alibaba that forms the basis of Qwen3.8-Max. It contains 2.4 trillion total parameters, with about 95 billion activated per token.

Where are the weights available?

The weights are available on Hugging Face, with Alibaba also providing access through ModelScope.

Is the open-weight model identical to Qwen3.8-Max?

No. The commercial Max version includes additional capabilities, including vision input, non-thinking support, a one-million-token context length by default and built-in tools.

Conclusion

The release of Qwen3.8-Max weights represents an important change in Alibaba’s strategy. For the first time, developers can access weights from the company’s Max-class Qwen lineup and deploy the 2.4-trillion-parameter model using standard inference frameworks.

The release also demonstrates the growing scale of China’s AI development. However, the downloadable model should not be confused with the full commercial Qwen3.8-Max service: the two versions differ in areas including multimodal input, context length and available features.

With strong benchmark results, a massive mixture-of-experts architecture and broader developer access, Qwen3.8-Max gives Alibaba another significant position in the increasingly competitive global open-weight AI market.