Artificial intelligence model distillation, as a key technology driving lightweight deployment of large models, serves as the core pathway for achieving “AI+” industrial integration. However, during its technological diffusion, it faces governance challenges stemming from inadequate intellectual property system adaptability. AI model distillation refers to the process of transferring knowledge from a pre-trained large teacher model to a lightweight student model through strategies such as softening the output distribution. This achieves model compression, enabling cost-effective and efficient deployment on terminal devices like smartphones and intelligent robots. Model distillation can be categorized into three approaches: soft label distillation, feature map distillation, and relational distillation. These correspond respectively to the distribution probability data output by the teacher model, the feature data input to the teacher model, and the inference process data from input to output within the teacher model. These data may potentially hold intellectual property rights. The distribution probability data output by the teacher model, the feature data input to it, and the inference process data from input to output may constitute works. If student model developers improperly replicate these data through model distillation, they may infringe copyright. If the aforementioned data constitute trade secrets, model distillation may infringe trade secrets by constituting improper means or violating confidentiality obligations. Distillation practices violating user agreements carry risks of breach of contract or unfair competition. Regarding these issues, international practice presents two typical models: the open innovation model, exemplified by DeepSeek, aims to achieve effective information utilization and sharing through open-source intellectual property initiatives, thereby reducing subsequent innovation costs. For copyright protection, it advocates automated, permissive open-source copyright licenses with licensed and restricted content; for trade secret protection, it proposes that open-source initiatives waive trade secret protection for much of the teacher model's data. Regarding unfair competition and breach of contract, it advocates for relatively lenient user agreements to completely exempt model distillation from breach of contract and unfair competition risks. The anti-competition model, exemplified by OpenAI, prohibits distillation through closed-source technology and strict agreements. Regarding copyright protection, it restricts other developers from directly appropriating the trainer model's data through closed-source models. It also supplements copyright protection by limiting other developers' use of model outputs for distillation via user agreements when copyright law rules are ambiguous. For trade secret protection, it safeguards the model's data as trade secrets through closed-source models and expands the scope of trade secret rights by restricting reverse engineering by other developers via user agreements. Regarding unfair competition and breach of contract, user agreements restrict model distillation by other developers, strengthening protection of their intangible competitive advantage—first-mover advantage. Open innovation pathways accelerate technological accessibility, foster innovation ecosystems, control industry standards and regulatory discourse, and counter external technological blockades and competitive pressures. Thus, they better align with China's industrial needs and scientific development patterns. China should cultivate an intellectual property-friendly industrial environment to promote open innovation models. It is recommended that, while respecting open-source agreements, China systematically resolve intellectual property conflicts in technology diffusion by clarifying copyright law rules for model distillation and defining the scope of anti-distillation agreements. This will foster a supportive institutional environment for AI technological innovation.