Kimi K3 open weights arrive with caveats: open access meets licensing thresholds for Model-as-a-Service

Kimi K3 open weights are now available for download, but the custom license introduces commercial thresholds that enterprises must evaluate before deploying the model in customer-facing products. The license permits broad use for internal purposes, but companies operating a Model as a Service business above $20 million in annual revenue must negotiate a separate agreement with Moonshot AI.

What happened

Moonshot AI released the full 2.8 trillion-parameter Kimi K3 weights, a Mixture-of-Experts model with a 1 million-token context window and 104 billion active parameters from 896 experts. Alongside the weights, the company published a 47-page technical report, optimized attention kernels, MoE communication libraries, and deployment components for self-hosting with vLLM and SGLang. The release makes Kimi K3 one of the largest open-weight models available, but the accompanying license is not a standard open-source license like Apache 2.0 or MIT.

Why AI builders should care

The Kimi K3 license grants broad rights to download, modify, and deploy the model for commercial purposes, but it imposes conditions on larger deployments. If you or your affiliates operate a Model as a Service business and aggregate revenue exceeds $20 million over any consecutive 12 months, you must enter a separate commercial agreement with Moonshot AI. The license also requires prominent attribution of "Kimi K3" on the user interface if your commercial product exceeds 100 million monthly active users or $20 million in monthly revenue. Internal use of the model is explicitly carved out from these requirements, making it viable for in-house research, developer tooling, and internal productivity workflows without triggering commercial obligations.

Practical implications

For enterprises, the first step is to determine whether your planned use qualifies as internal or external. Internal use (not making the model's capabilities available to third parties) avoids the commercial licensing and attribution requirements entirely. For external deployments, assess whether your offering constitutes Model as a Service under the license definition: giving third parties access to inference or fine-tuning via API with meaningful control over inputs, parameters, or training data. If it does, and your company or its affiliates exceed the $20 million annual revenue threshold, you must negotiate a separate commercial agreement. Additionally, if your product reaches 100 million MAU or $20 million monthly revenue, you must display "Kimi K3" prominently on the user interface. The revenue calculation includes aggregate revenue of the licensee and its affiliates, not just revenue from products built on Kimi K3.

Caveats

Open weights are not the same as open source. The Kimi K3 license is a custom permissive license with commercial thresholds, similar to Meta's Llama community license but with different mechanisms. The model is large: roughly 1.5 TB of weights, requiring significant infrastructure to run. While reports indicate successful deployments on clusters of consumer RTX 5090 GPUs, most organizations will need datacenter-grade hardware. The license also does not limit the revenue calculation to products built on Kimi K3, potentially bringing smaller subsidiaries of larger parent companies under the commercial terms. Enterprises should evaluate licensing terms with the same rigor as benchmarks and security reviews before deploying Kimi K3 in production.