moonshotai/Kimi-K3

Context: Why talk about Kimi-K3 now?

This July, Moonshot released the weights for their flagship model Kimi-K3 as promised. With 2.8 trillion parameters and weight files totaling 1.56TB, it's already available from 7 providers on OpenRouter at the same pricing as Moonshot's own API ($3/M input, $15/M output). This is a significant industry event, showing top Chinese AI companies actively joining the global open ecosystem. But more intriguing is the attached license, which reveals a fundamental shift in what 'open source' means in the era of large language models.

Breakdown: What exactly does Kimi-K3's license say?

Let's trace the changes from K2 to K3. When Kimi-K2 launched in July 2025, Moonshot used a 'modified MIT license.' The core change was simple: if your commercial product has over 100 million monthly active users or over $20 million monthly revenue, you need to prominently display 'Kimi K2' in the UI. It was an attribution requirement with a high threshold, affecting a limited scope.

But with K3, the nature of the license changed. First, it no longer calls itself 'modified MIT'—it's simply the 'Kimi-K3 License.' Second, a key clause was added: if you or your affiliates operate a 'Model as a Service' (MaaS) business, and your aggregate revenue exceeds $20 million over any consecutive 12 months, you must enter into a separate agreement with Moonshot before using the Software or its derivatives for any commercial purpose.

This is no longer just an attribution request; it's a substantial commercial licensing barrier. Imagine you're a fast-growing AI startup using K3 as your base model for API services. As your annual revenue approaches that $20 million line, you face a choice: either negotiate a commercial deal with Moonshot, or stop using K3. This effectively transforms K3 from a freely commercializable 'open source' tool into an 'open weight' resource with specific conditions.

Trend Insight: The divide between 'open source' and 'open weight' is solidifying

Moonshot has been very honest here. They consistently use the term 'open weight' rather than 'open source' in their materials. This reflects a deeper industry trend: pure open source, as defined by OSI (Open Source Initiative), is becoming increasingly rare in the large model space.

Why? Because training a frontier model is extraordinarily expensive (tens to hundreds of millions of dollars). If fully open-sourced, companies like Meta, Google, or any large player could directly commercialize it, potentially fueling their own model services without sharing revenue with the original creators. This poses huge commercial risks for companies investing heavily in R&D.

Thus, we see the rise of a new paradigm: 'open weights + commercial terms.' Meta's Llama series (with monthly active user limits), some Mistral models, and now Kimi-K3 all follow this pattern. They open up the core capability (weights), allowing research and small-to-medium scale commercial use, but set up 'toll gates' for large-scale commercial services. This strikes a pragmatic balance—promoting technology adoption while protecting developers' business interests.

Practical Value: What does this mean for developers/businesses?

1. Always read the license carefully before choosing a model. 'Open source' and 'open weight' are no longer synonyms. You need to clarify whether your business model might hit the revenue or user threshold.
2. Assess business risks. If your business is providing model APIs (MaaS) or expects rapid growth in users/revenue, using models like Kimi-K3 carries uncertainty about future renegotiations. Truly permissive licenses like Llama's, Pythia, or certain fully open-source models might be 'safer' choices, even if their capabilities differ.
3. Monitor pricing. Kimi-K3's pricing via OpenRouter is highly competitive. Even if you don't use its weights, calling its API is a cost-effective option, which itself is driving price competition in the inference service market.

Counterintuitive Insight: License terms are becoming part of product design

Many may not realize that license terms are no longer just legal appendixes; they're central to product strategy and ecosystem building. Through K3's license, Moonshot is effectively constructing a tiered ecosystem: small developers can use it freely, fostering innovation and community growth; large MaaS players must become business partners. This resembles the freemium model seen in early game engines or cloud platforms. Licensing terms are becoming a key lever for building moats and business models in the AI era.