Alibaba says Qwen3.8-Max ran 33+ self-improvement cycles, lifting its Artificial Analysis score to 45

Alibaba is making one of the boldest claims of the week. According to HPCwire's coverage of Chinese AI progress, Qwen3.8-Max went through more than 33 cycles of recursive self-improvement, meaning the model helped generate improvements to itself. That process reportedly raised its Artificial Analysis intelligence score from 40 to 45. A five-point gain on that aggregate index is meaningful, and Alibaba credits it to the self-improvement loop rather than a new pretraining run.

On the open side, Alibaba's Qwen team says Qwen3.8-27B is now available through Nebius Token Factory. It is pitched as a dense model for agent workflows and multi-step research. Mid-size dense models like this are popular with teams that self-host or fine-tune.

The Qwen story is moving onto devices as well. Alibaba launched Qwen Intelligence, an on-device agent framework debuting on HONOR's Magic9 flagship phones. It includes Mobile Planner, Mobile-Use and Mobile Creative agents that can run tasks across apps. That competes directly with Apple's Siri AI and Google's Gemini on Android.

Competitive context: Chinese labs are gaining on several fronts at once:

- DeepSeek leads OpenRouter by volume.

- DeepSeek and Huawei are open-sourcing Ascend chip tooling.

- Alibaba is claiming self-improvement gains.

The claims deserve skepticism. "Recursive self-improvement" is a loaded term, and the sources do not say what each cycle involved, how much human oversight there was, or whether the gains hold on benchmarks the model was not tuned toward. Independent replication on Artificial Analysis and on agentic evaluations will be the real test.