Breaking

Alibaba’s Qwen3.8-Max Completes Multi-Day Autonomous Coding, Beats a Paper by 2.71 Points, and Outranks 458 Contest Teams

September 5, 2026 at 22:45 EDT

Foundation Models
AI Agents
Software Dev & Coding

Alibaba’s Qwen team says Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters, completed multi-day coding and research runs without a human in the loop, then placed ahead of 458 of 526 teams in an online machine-learning contest. The flagship, available on QwenCloud with a planned open-weight release after launch, is positioned for long-horizon software, research, and agent work rather than one-shot chat.

August 2026 · Alibaba Qwen

Unattended for days, then a 2.71-point paper win

Qwen3.8-Max ran a 10-plus-day coding project, spent ~125 hours reproducing then beating a research paper, and finished ahead of 458 of 526 human teams in a 24-hour ML contest.

10+

Days of unattended coding on oh-my-cli

7,600

Lines of code in the ~125-hour paper run

+2.71

AIME24 points vs reproduced paper method

458

Of 526 contest teams placed behind it

AIME24 after self-improvement

Same unit: percent solved. Column height scales with the reported score.

49.58%

Paper method, reproduced

52.29%

nhighgate method (round 4)

Takeaway: four hypothesis rounds moved AIME24 by 2.71 points over the paper baseline.

Research loop

1. Paper + GPUs
No starter code

→

2. Reproduce
~37 hours, six findings

→

3. Iterate
18 ideas, 4 rounds

→

4. Beat paper
+2.71 on AIME24

What held up

Multi-day traces, a public GitHub log, PaperBench 93.0, and a real contest climb from 0.60 to 0.853 accuracy.

What is still open

Independent bug-fixing lagged GPT-5.6 Sol (19/105 vs 42). SWE-bench Pro trails Fable 5 (67.7 vs 80.0). Third-party Terminal-Bench scores are lower than Alibaba’s table.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.