Alibaba released Qwen 3.8-Max, its largest open model: 2.4 trillion total parameters, 95 billion active per query thanks to Mixture of Experts architecture. Weights will be downloadable next week. The model is already accessible via API on Alibaba platforms (Qoder, QoderWork, Token Plan).

The Qwen team claims performance on par with Anthropic’s Claude Fable 5. On the crowdsourced Arena.AI leaderboard, Qwen 3.8-Max ranks behind only Fable 5 and three Opus-family models. For frontend coding, only two Opus and Moonshot’s Kimi K3 rank ahead.

Why this matters to you. Six days ago, Kimi K3 from Moonshot became the first open Chinese model served on US GPU infrastructure, via Telnyx API with OpenAI-compatible endpoints. Now Alibaba takes a similar step with a larger model. Chinese open models are shifting from downloadable weights for researchers to concrete options for product builders.

If your stack relies on OpenAI or Anthropic APIs, a model at this capability level changes the calculation. You can evaluate alternatives to the American frontier without waiting, and single-provider dependence shrinks.

Alibaba’s use cases go beyond chat: in sixteen days the model built a command-line tool by itself (265 commits, 127 pull requests, zero human intervention). It reproduced research paper results and improved them by 2.7 points on AIME24. It managed a simulated year of e-commerce and quadrupled starting capital.

These numbers come from Alibaba’s internal testing: treat them cautiously until independently replicated.

If you want to try it: the model is already available on the Alibaba platforms mentioned. Weights for self-hosting arrive next week.

In depth

Alibaba returns to open weights for its advanced models after exploring the proprietary route. Qwen 3.8-Max is the first model in the Max class (its most capable tier) whose weights will be publicly downloadable. The move comes six days after Moonshot released Kimi K3 (2.8T parameters) and a week after Kimi K3 entered US production stacks via Telnyx, with OpenAI-compatible endpoints. Two Chinese giants, two converging paths toward the same goal: making open models an operational product, not just a research artifact.

Architecture in brief. Qwen 3.8-Max is a Mixture of Experts: instead of activating all 2.4 trillion parameters for each response, it uses 95 billion per query, selecting the most suitable “experts” for each task. The result is a model with a giant’s theoretical capacity but inference costs closer to a 95-billion-parameter model. For non-model-trainers, the practical translation is: you can have frontier performance without paying frontier rates per token.

What the use cases show. The Qwen team published five demonstrations that go beyond static benchmarks. The model built a CLI tool in sixteen days of autonomous work (265 commits, 127 pull requests). It reproduced research paper results, Unified Data Selection for LLM Reasoning, in roughly 125 hours of GPU compute, writing 7,600 lines of code and beating the original method by 2.7 points on the AIME24 benchmark. It managed a simulated fiscal year on Taobao and Tmall data, starting with 100,000 yuan and ending with 416,252, quadrupling initial capital and beating the second-place finisher (GLM 5.2) by 38%.

These are weighty numbers, but all come from Alibaba’s team. Their purpose is demonstrating long-horizon planning capability (tasks spanning days, not single chat turns), but without independent replication they remain company claims. The Arena.AI leaderboard, which is crowdsourced and not controlled by Alibaba, provides partial corroboration: Qwen 3.8-Max ranks fifth overall, behind Fable 5 and three Opus models.

Where sources diverge. The Verge and The Decoder agree on the 2.4 trillion parameters and Fable 5 comparison. The Decoder adds details on 95 billion active parameters and the five use cases; The Verge focuses on geopolitical context and the leaderboard. Neither source explicitly confirms partnerships with US-based inference providers: operational details on who will serve the model on US soil remain unverified.

The limits. Alibaba announced weights for next week but hasn’t yet published the specific license. “Open-weight” in the Chinese context means downloadable weights with usage restrictions, not open source in the classical sense: experience with Kimi K3 suggests waiting for concrete terms before assuming the model is free for commercial use.

The regulatory context is in motion. The White House is pursuing selective bans on specific Chinese open-weight models, and OpenAI and Anthropic are making common cause on open-weights risks. If Chinese open models keep converging toward American frontier capacity, political pressure to limit their adoption in US enterprise stacks will grow. For product builders, the window for testing these alternatives without regulatory constraints may not stay open long.