Alibaba Plans Revenue Share for Qwen3.8-Max

The era of free open-weight frontier models from China may be ending. Alibaba plans to require large commercial users of its 2.4-trillion-parameter Qwen3.8-Max model to share a portion of revenue, Reuters reported Thursday, citing unnamed sources — making it the second major Chinese AI lab in two weeks to attach commercial terms to open-weight releases, following Moonshot's Kimi K3.

What happened

Alibaba reportedly plans to announce revenue-sharing terms for Qwen3.8-Max as early as next week, with open weights expected around August 10, according to Reuters' sources. The exact percentage and the threshold for "large commercial users" have not been published.

Until now, Alibaba permitted commercial deployments of Qwen models on customers' own infrastructure without charge. The Apache 2.0 license for Qwen3 open-weight models carried no revenue-sharing condition. Qwen3.8-Max — a 2.4-trillion-parameter MoE model currently in preview — would be the first Qwen model with commercial strings attached.

Moonshot's K3 established the template: up to 30% revenue share for MaaS providers exceeding $20 million in annual revenue. Alibaba's terms, according to Reuters' sources, may follow a similar structure — though the company has not confirmed any specifics.

Why it matters

This isn't a one-off. Two of China's three leading AI labs are now moving toward the same business model for open-weight frontier models: free for research and small-scale experimentation, revenue-sharing for commercial-scale deployment. The third — DeepSeek — has yet to signal its approach, but the direction is clear from Reuters' reporting.

Alibaba's Qwen ecosystem is the largest in open-weight AI by derivatives on Hugging Face. Those derivatives were built under Apache 2.0 with no commercial strings. Revenue-sharing terms would segment that ecosystem into two tiers — free for experimentation, paid for production at scale — and introduce a new cost variable for any enterprise evaluating Qwen3.8-Max as a self-hosted alternative to API-gated models from OpenAI or Anthropic.

The shift also redefines what "open-weight" means commercially. The weights are downloadable, but deploying them at scale now carries a licensing cost. It's a middle ground between fully open (Apache 2.0, no restrictions) and fully proprietary (API-only).

What changes for you

If you're evaluating Qwen3.8-Max for production: Budget for potential revenue-sharing terms. The Moonshot K3 precedent (30% of MaaS revenue above $20M) gives a rough ceiling — Alibaba's terms are unlikely to be stricter than a direct competitor's.

If you're a researcher or small-scale user: Nothing changes yet. Both Alibaba and Moonshot have indicated thresholds apply only to large commercial deployments. Experimentation and academic use remain free.

If you're building on the Qwen ecosystem: Monitor the license terms closely when they drop next week. The Apache 2.0 licensing for existing Qwen3 models is unchanged — only Qwen3.8-Max will carry the new terms at launch.

FAQ

Q: Will this affect existing Qwen models?
No. The revenue-sharing terms apply only to Qwen3.8-Max, not to previously released Qwen models under Apache 2.0.

Q: How does this compare to Meta's Llama open-weight approach?
Meta's Llama models have a custom license that restricts use by large-scale competitors but does not require revenue sharing. Alibaba and Moonshot are pioneering a different model — open weights with a direct commercial royalty.

Q: Is DeepSeek likely to follow?
DeepSeek has not announced any revenue-sharing plans, but with both Moonshot and Alibaba adopting the model, pressure is mounting. DeepSeek's V4 models are currently API-only; any open-weight release would now face the question of whether to attach commercial terms.