Full Breakdown
Alibaba Unveils Qwen3.8-Max, Its Largest Open-Weight AI Model
8/4/2026, 1:54:03 AM
Core Event
- On August 3, Alibaba Group Holding announced the launch of Qwen3.8-Max, calling it the most capable model in its Qwen series.
- The model contains 2.4 trillion parameters, supports text, image and video inputs, and offers a context window of up to 1 million tokens.
- It uses a mixture-of-experts design that activates roughly 95 billion parameters per request, a strategy intended to cut compute cost and latency.
- Alibaba reported that the model completed an autonomous software-engineering project in 16 days.
- The company plans to make the model’s weights publicly downloadable in the week starting August 10.
Background & Context
- Chinese AI firms have accelerated open-weight releases to compete with U.S. labs such as OpenAI, Anthropic and Google, which keep model weights private.
- Moonshot AI’s Kimi K3 (released in July with 2.8 trillion parameters) set a recent benchmark that Qwen3.8-Max closely follows.
- Beijing promotes open-weight models as a way to expand global influence in AI governance and to lower barriers for developers worldwide.
Data & Statistics
- Parameters: 2.4 trillion total; 95 billion active per inference.
- Token window: up to 1 million tokens.
- Arena.AI rankings: first among Chinese text models; second globally for vision tasks, behind Anthropic’s Claude Fable 5 variant.
- Stock reaction: Alibaba shares rose ?4.5 % in pre-market trading in New York and ?7 % on the Hong Kong exchange on August 3.
- Pricing: US$2 per million input tokens and US$6 per million output tokens.
Official Statements & Responses
- U.S. Treasury officials, cited in related reporting, indicated ongoing scrutiny of Chinese AI models for possible export-control violations, though no specific allegation was made against Qwen3.8-Max.
Why It Matters
- The open-weight release positions Alibaba to attract a global developer community and could boost demand for Alibaba Cloud services.
- Activating only a fraction of its parameters lowers operating costs, potentially intensifying price competition with U.S. providers whose models remain closed.
Conflicting Reports & Gaps
- Alibaba’s claim that Qwen3.8-Max ranks first among Chinese text models and second globally for vision tasks is based on the company’s own submission to Arena.AI; independent verification has not been published.
- Some sources say the model “sometimes exceeds” Anthropic’s Claude Fable 5 on internal benchmarks, while others note it still lags behind several Anthropic Opus variants. The absence of third-party benchmark results leaves the true comparative performance uncertain.
Verbatim Quotes
- “Many investors continue to underestimate Chinese AI models because of US chip restrictions or general scepticism,” — Vey-Sern Ling, of Qwen3
What's Next
- The model’s weights are scheduled for public release in the week beginning August 10, after which developers can download them from platforms such as Hugging Face and ModelScope. Alibaba has indicated that the open-weight version will be integrated into its QwenWork productivity suite and made accessible via standard OpenAI-compatible API endpoints.
