Behind the Western assumption that Chinese tech firms operate as a unified open-source front against proprietary American labs, a sharp commercial divide has emerged. An in-depth investigation by 36Kr reveals that Chinese foundation model developers have split into two opposing licensing camps: incumbents protecting their cloud moats through restrictive commercial agreements, and challenger labs doubling down on pure open source.

The friction centers on Model-as-a-Service (MaaS) redistribution: whether rival hyperscalers can monetize open weights without paying the labs that funded their pretraining.

Confirmed

  • Alibaba Qwen commercial gates: Alibaba’s Qwen3.8-Max and experimental Qwen3.8-Flash-Next were released under the Qwen Community License 1.0. The license mandates formal commercial agreements for companies with over $50 million in annual revenue and explicitly restricts third parties from offering MaaS or AI work assistant platforms built on Qwen without separate licensing.
  • Moonshot AI (Kimi K3): Kimi K3 introduced equivalent commercial thresholds, restricting unapproved cloud hosting and aggregating corporate affiliate revenue to enforce enterprise licensing requirements.
  • DeepSeek stays MIT: DeepSeek reaffirmed its commitment to unconstrained open source by releasing DeepSeek V4-Pro under the standard MIT License, imposing zero revenue thresholds, registration requirements, or commercial redistribution bans.
  • Zhipu AI (GLM-5.3-Flash): Zhipu continued its transition toward full permissiveness, releasing GLM-5.3-Flash under standard MIT terms without commercial registration gates—a departure from its earlier ChatGLM-2/3 era restrictions.
  • Market blast radius: 36Kr reports that independent token aggregators such as SiliconFlow (reporting ~$7.7 million in annual revenue) fall well beneath the $50 million threshold. The restrictions are engineered primarily against domestic cloud giants: ByteDance’s Volcengine, Baidu Intelligent Cloud, and Tencent Cloud.

Unknown

  • Cross-border enforceability: How Alibaba or Moonshot intend to enforce anti-MaaS restrictions against Western or European cloud aggregators reselling Qwen weights outside mainland jurisdiction.
  • Impact on open distillation: Whether downstream developers fine-tuning Qwen models for SaaS wrappers risk license violations if their parent companies surpass revenue thresholds.
  • SiliconFlow IPO dynamic: How shifting license terms will impact SiliconFlow’s pending Hong Kong listing, given its core business depends on serving multi-provider open weights.

Why it matters

For global engineering teams, "open-weight" is no longer synonymous with "license-safe." A startup building on DeepSeek V4 or GLM-5.3 operates under standard MIT indemnity, whereas building an enterprise inference gateway on Qwen3.8-Max requires tracking corporate revenue ceilings and navigating explicit anti-MaaS boundaries. The move proves that open weights in China are increasingly weaponized as cloud loss-leaders: free for end-users, but strictly fenced against rival cloud hosts.

Our take

The Chinese open model ecosystem is maturing out of its market-share land-grab phase into classic enterprise protectionism. When models like Qwen reached frontier parity, Alibaba realized it was subsidizing competitor infrastructure on ByteDance and Baidu clouds. Expect the divide to widen: tier-one cloud providers will increasingly embrace synthetic "community" licenses with cloud carve-outs, while independent labs like DeepSeek use permissive MIT licenses to commoditize everyone else’s infrastructure.

Sources