Qwen3.8-Max recursive self-improvement demos (Apsara 2026)
Alibaba says Qwen3.8-Max ran 33 automated self-improvement cycles in a month (Artificial Analysis score 40 to 45) and cut a chip module's area 42%.
Company-reported demos of the model improving itself. Over a month of fully automated runs, 33 cycles lifted its Artificial Analysis score from 40 to 45, and in a chip-design experiment, 60+ hours and 10,000+ EDA tool calls produced bus modules with 42% less area and no performance loss. Methods and baselines were not published.
- Date
- Tuesday, 22 September 2026
- Lab
- Alibaba (Qwen)
- Kind
- feature
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Iterative cycles / Artificial Analysis score | 33 cycles; 40 to 45 Over about a month of fully automated runs, per Alibaba's recap | company |
| Chip-area reduction | 42% 60+ hours, 10,000+ EDA tool calls, no performance loss (company-reported) | company |
The dates conflict. Alibaba Group's Apsara recap is dated 2026-09-25 and its header gives that day for the conference, while other coverage places the Qwen 4 announcement on 2026-09-22; the later date is not ruled out. These are company claims from a keynote recap, with no named experimenter, paper or independent replication; whether the Artificial Analysis score is its Intelligence Index is not stated.
Sources
- www.alibabagroup.com/en-US/document-2041268239081668608
- rits.shanghai.nyu.edu/ai/alibaba-at-apsara-2026-qwen-4-in-training-10t-models-next/
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check