Alibaba's most powerful model ever is going open source

Alibaba's most capable model line used to be the one you couldn't have. The Qwen-Max series was built for paying cloud customers and stayed closed even as smaller open-weight Qwen models spread widely. That wall just came down. Alibaba announced Qwen3.8-Max, which the company calls the most capable model in its family to date, in a post on its cloud blog. A 2.4-trillion-parameter model built on the architecture of Qwen 3.5, it is the first Max-class flagship to be open-sourced: the weights land next week on Hugging Face and ModelScope.

The release post rests on three long autonomous runs, each with a public trace, in which the model improved its own work over days without human help.

Open weights at the top of the line

Open-sourcing a flagship is a move few labs choose, because keeping the best weights behind a paid API is how frontier AI companies protect their margin. Alibaba is pairing the release with a product push: the company has opened the beta of QwenWork, a workplace agent platform that folds desktop, cloud, and the DingTalk collaboration suite, which serves more than 20 million enterprises and organizations, into one subscription. It is the platform-first bet Alibaba is making for the second half of 2026: lock in the enterprise before the model race is over. Qwen3.8-Max is a highlighted model option on QwenWork starting Aug 3, and the platform's four model tiers run from Economy to Flagship.

For developers, the new model speaks the API dialects of its rivals. QwenCloud offers OpenAI-compatible chat completions and responses endpoints plus an Anthropic-compatible interface, so Qwen3.8-Max can run inside Claude Code or Codex with a couple of environment variables, the same compatibility that lets the Qwen Cloud Token Plan work inside Claude Code. A reasoning_effort parameter, with xhigh, medium, and low levels, controls how much thinking the model does before it answers, and preserve_thinking is on by default.

Four rounds that beat the paper

The most detailed evidence is the research-reproduction run. Handed a paper called "Unified Data Selection for LLM Reasoning" and no starter code, Qwen3.8-Max spent roughly 37 hours rebuilding the pipeline from zero, wrote about 7,600 lines of code, ran 33 rounds of GPU training, and reproduced the paper's six main findings. Then it used another 88 hours to try to do better, generating 18 ideas across four rounds. The winning idea, counting what the model calls hard decision points in training examples, beat the paper's own method by 2.7 points on the AIME24 math benchmark.

The other two runs follow the same pattern. Entered into the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba's Tianchi platform, the model fine-tuned several language models, merged them into a weighted voting system, and climbed from 0.60 to 0.853 accuracy across 45 submissions, finishing ahead of 458 of the 526 human teams, a result we reported when the challenge ended. In a chip-design exercise, it took a GCD/RSA accelerator from 8,298 gates to 678 gates over roughly 500 turns, and the physical layout shrank from 106×106 µm² to 46×46 µm². A 16-day autonomous run on the oh-my-cli project accumulated 265 commits, 127 PRs, and 151 issues, all visible in the public GitHub repository.

The house scoring its own game

None of this comes with an independent auditor. The results are Alibaba's own, and the most striking ones sit on benchmarks Alibaba designed. On E-Commerce Bench, a 365-day business simulation built on desensitized Taobao and Tmall transaction data, Qwen3.8-Max turned ¥100,000 of starting capital into a ¥416,252 balance, a 4.16x return that the company says beats second-place GLM 5.2 by 38%. The harness comparisons against Claude Code and Codex come from the same blog. None of that makes the results false. It means the verification burden falls on everyone else, at a moment when the static tests that used to arbitrate these claims are themselves collapsing.

The open weights solve part of that problem. Once they land next week, anyone with enough GPUs can rerun the contested benchmarks and check whether the model holds up outside Alibaba's own scaffolding. The frontier labs keep their biggest weights behind paid APIs; Alibaba just gave its best model away. The verdict on Qwen3.8-Max will come from the open-source community, not from a launch post.