Alibaba releases Qwen3.8-Max, will open-source its weights
Alibaba's Qwen team released Qwen3.8-Max, a 2.4-trillion-parameter model it calls its most capable yet, and said it will publish the weights next week, the first open release of a Max-class Qwen model.
Alibaba's Qwen team announced Qwen3.8-Max, which it calls the most capable model in the Qwen family. The team said the open weights will follow next week. It is the first time a Qwen-Max-class model will be open-sourced. The model builds on the Qwen 3.5 architecture and scales to 2.4 trillion parameters. Qwen says it improves on earlier models in coding, work, research and long-horizon tasks.
Most of the announcement covers long autonomous agent runs:
- Self-evolving coding harness. Qwen says the model built the
oh-my-cliproject from scratch in a run lasting more than 10 days. The run turned GitHub issues into tasks that agents claimed, implemented, tested and merged. After about 16 days of autonomous operation, the repository had 265 commits, 127 PRs and 151 issues as of July 30, 2026. The full trace is public atqwen-code-dev-bot/oh-my-cli. - Reproducing a paper. Starting from the paper "Unified Data Selection for LLM Reasoning" and a set of GPUs, the model worked for about 125 hours. It wrote roughly 7,600 lines of code and ran 33 GPU training rounds. Qwen says it reproduced the paper's six main findings by fine-tuning Qwen3-8B. Over four rounds it then tested 18 of its own ideas. The best one raised AIME24 from the reproduced baseline of 49.58% to 52.29%, a gain of 2.71 points.
- Kaggle-style competition. Qwen also entered the model in the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba Cloud's Tianchi platform, which had 526 human teams. The available excerpt cuts off before the result.
The excerpt includes no standard benchmark tables, pricing, API details or license terms for the coming weights. The demonstrations are Qwen's own reports and have not been independently checked.