In brief
- The Qwen3.8-Max model features 2.4 trillion parameters and is the first Max-class Qwen model released with open weights by Alibaba.
- It includes setup instructions for competing tools, Claude Code by Anthropic and Codex by OpenAI.
- According to Alibaba's benchmarks, Fable 5 excels in 15 out of 31 tests, while Qwen3.8-Max secures victories in seven.
On Monday, Alibaba unveiled its Qwen3.8-Max model, which it claims to be its most advanced AI model to date. This model will be available on Hugging Face and ModelScope next week, marking the first time Alibaba has released a Max-scale model for free.
This model boasts impressive specifications, with a total of 2.4 trillion parameters, of which 95 billion are activated at any given time. These parameters represent the various adjustable aspects of the model.
This efficiency is crucial, as it allows the model to operate without excessive resource demands. One can envision it as a vast library where only the relevant sections illuminate in response to specific inquiries. Consequently, small enterprises and research labs with adequate hardware can utilize a cutting-edge model without incurring the costs associated with large data centers.
Alibaba's announcement shifts focus from the typical benchmark competitiveness to the model's endurance. It autonomously developed a coding tool over 16 days, achieving 265 commits, 127 pull requests, and addressing 151 issues without any human intervention. Additionally, it took five days to replicate a research paper's coding without prior exposure, ultimately surpassing the original results by 2.7 points. In a 24-hour machine learning competition, it outperformed 458 out of 526 human teams.
Designed for Compatibility with Rivals' Tools
Qwen3.8-Max comes with guidance for both Claude Code and Codex, developed by Anthropic and OpenAI, respectively. Alibaba's API is compatible with the protocols of both companies. Most coding benchmarks were conducted using Claude Code.
However, the results from these benchmarks are not particularly flattering. In a series of 31 text-based tests, Anthropic's Fable 5 claimed first place in 15 instances, while OpenAI's GPT-5.6 Sol achieved nine victories and Qwen3.8-Max won seven. In the realm of coding assessments, Qwen managed to secure only one win. Nevertheless, in terms of operational costs, this model is remarkably economical, making it nearly 30% cheaper than Claude Fable 5, despite possibly requiring more iterations or reasoning to complete tasks.
In contrast, when it comes to multimodal tasks—such as handling documents, video, and spatial reasoning—Qwen3.8-Max leads in most categories.
There has also been a strategic shift for Alibaba. Back in April, the company discontinued the free tier of Qwen Code, moving towards more closed, paid models following leadership changes. Our previous review of Qwen 3.7 Max indicated that while the Plus version would remain open, the Max version would be restricted to API access.
Now, that policy has changed, and the timing appears deliberate. The share of Chinese open-weight models on OpenRouter surged from less than 2% in late 2024 to approximately 61% by mid-2026. Qwen has now surpassed Meta's Llama as the leading self-hosted model globally.
Meanwhile, Washington imposed export controls on Fable 5 and Mythos 5 in June, and reports suggest that Beijing is contemplating its own restrictions on Chinese models being exported abroad.
Thus, while Alibaba may be trailing in some benchmarks, it is excelling in terms of distribution. Offering a comparable model for free positions it favorably in the market, even if it is not the top performer.