Chinese AI startup Moonshot AI has released the full weights for Kimi K3, its largest model to date. The 2.8 trillion-parameter Mixture-of-Experts system comes with a one-million-token context window and strong benchmark results, along with inference code, optimized kernels, and support for vLLM and SGLang.
The accompanying license allows broad use, modification, and commercial deployment. However, two clauses create specific obligations once organizations reach certain revenue or user thresholds. Companies operating a Model-as-a-Service offering that exceeds $20 million in annual revenue must negotiate a separate commercial agreement. In addition, products or services surpassing 100 million monthly active users or $20 million in monthly revenue must display the name “Kimi K3” prominently in the interface.
Internal deployments remain largely unrestricted. Any use that keeps model outputs inside the organization avoids the revenue-based requirements. This distinction matters for banks, manufacturers, and service firms that plan to run the model on private infrastructure for employee tools rather than customer-facing APIs.
The license structure reflects a growing pattern among frontier labs. Meta’s Llama models impose a user threshold before a commercial deal is required, while other developers attach redistribution or attribution rules. Moonshot’s approach ties obligations to overall company revenue rather than model-specific usage, which can pull affiliates of larger parents into the same compliance category.
One practical consequence is that smaller AI startups offering fine-tuning APIs may need to budget for licensing fees once they cross the $20 million mark, even if their Kimi K3 revenue is modest. Conversely, enterprises that embed the model only inside internal productivity suites face fewer hurdles. This clarity could speed adoption among non-tech verticals that have hesitated to adopt open-weight frontier models because of ambiguous legal terms.
Another implication concerns supply-chain risk. Organizations evaluating Kimi K3 for regulated workloads must still conduct security and compliance reviews, yet the availability of full weights and training documentation gives them more visibility than API-only access. At the same time, the 1.5 TB model size means only well-resourced teams can run it without substantial infrastructure investment.
Developers have welcomed the accompanying technical report and ecosystem integrations. Early reports of successful inference on consumer-grade GPU clusters suggest the release may broaden experimentation beyond hyperscale data centers. Still, most production deployments will likely remain with organizations that already operate large GPU fleets.
Enterprise leaders should map their intended usage against the license before downloading weights. Teams planning customer-facing AI services need to model potential revenue growth against the $20 million threshold and prepare for possible attribution or separate licensing steps. Those restricting the model to internal tools can proceed with greater flexibility. The release therefore underscores that open-weight availability is only one part of the evaluation; the attached legal framework now requires the same level of scrutiny as performance benchmarks or infrastructure costs.
