Qwen3.8-Max: A New Standard for Coding and Collaboration

Qwen has unveiled its latest Max-tier model, Qwen3.8-Max. Built on the Qwen 3.5 architecture, it uses a total of 2.4T parameters with 95B active parameters, and strengthens performance and stable end-to-end task execution across coding, work, research, and long-horizon tasks. The model is currently available via the QwenCloud API, with open weights set to be released next week.

The core focus of this evaluation is not simple code generation, but the ability to complete an entire real-world project from start to finish without human intervention. Rather than following a fixed plan, Qwen3.8-Max was tested on its ability to improve its own working methods through feedback loops.

In an autonomous coding experiment lasting over 10 days, it built the oh-my-cli project from scratch, creating a self-evolving harness that automated everything from requirements gathering, issue creation and assignment, code writing, testing, previewing, log analysis, and bug fixing.

• It managed tasks using an issue state machine based on GitHub Issues, combined with a dispatcher, monitor, and watchdog.
• After each change, it ran Build, Unit Test, E2E, and Desktop Lifecycle verification, and when anomalies occurred, it rolled back to related issues and PRs for re-verification.
• It converted community and user feedback into actionable tasks, continuously improving features such as /goal, /resume, Dynamic Workflow, Session Replay, and Desktop.

As of July 30, 2026, the repository—operated fully autonomously for about 16 days—accumulated 265 commits, 127 PRs, and 151 issues.

In another experiment, based on the research paper "Unified Data Selection for LLM Reasoning," it reproduced the entire data processing, training, and evaluation pipeline from scratch and even attempted performance improvements. Over about 125 hours, it wrote 7,600 lines of code, performed more than 1,100 tasks and 33 GPU training runs, and spent about 37 hours rebuilding the paper's full pipeline.