Kimi K3 weights are out: the real cost is in the license, not the benchmarks

Moonshot AI has released the full weights for Kimi K3, completing the rollout of what the company describes as the world's first open 3T-class model. The technical specifications are substantial: a 2.8 trillion-parameter Mixture-of-Experts architecture that activates 104 billion parameters from a pool of 896 experts, a one million-token context window, and a 47-page technical report documenting training innovations including Kimi Delta Attention, Attention Residuals, and Stable LatentMoE. Inference support arrived quickly through vLLM and SGLang, and the approximately 1.5 TB of weights have already been tested on consumer GPU clusters. But the publication that matters alongside the benchmark charts is the Kimi K3 License itself.

The license grants broad rights to download, modify, and deploy the model for commercial purposes, which represents a genuine expansion of access for organizations seeking frontier-level AI they can control and run offline. The structure, however, introduces commercial obligations that are not present in traditional open-source licenses like Apache 2.0 or MIT, and the way those obligations are scoped deserves close attention from any enterprise considering self-hosting.

Section 2 of the license is the provision that drew immediate attention from researchers and developers. Organizations operating a "Model as a Service" business, defined as giving third parties meaningful control over inputs, parameters, or training data through API access, must enter a separate commercial agreement with Moonshot AI if their aggregate revenue exceeds $20 million over any consecutive twelve months. The definition explicitly carves out end-user products with embedded model capabilities and mere request relaying, which suggests that consumer brands using Kimi K3 as a customer service chatbot are not captured. The definition is narrow enough to exclude many common enterprise deployments, but it is broad enough to catch infrastructure providers, AI service platforms, and startups building model-centric tooling.

The revenue threshold is calculated against the licensee's aggregate revenue and that of its affiliates, not against revenue generated specifically by Kimi K3-based products. This distinction matters for legal and finance teams at organizations with multiple subsidiaries or business units. A smaller company that is part of a larger parent generating $20 million annually falls under the commercial licensing requirement if it uses Kimi K3 in a Model-as-a-Service capacity anywhere in the group structure. That framing departs from the typical approach in commercial AI licensing, which ties obligations to revenue from the specific product rather than the broader company.

Section 3 adds a separate layer: commercial deployments reaching more than 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3" prominently in the product interface. For enterprise software vendors, AI copilots, and consumer applications that typically abstract away the underlying model, this requirement introduces a branding and disclosure obligation that does not exist under standard open-source licensing. The practical effect is that products built on Kimi K3 at scale must identify the model to end users, which may conflict with existing contractual commitments, white-label arrangements, or product strategies that treat model disclosure as an implementation detail rather than a user-facing disclosure.

Internal use receives a meaningful carve-out. The license exempts deployments that do not make the model, its outputs, or its underlying capabilities available to third parties, which means development teams, legal departments, and research groups running Kimi K3 for internal workflows are not subject to Sections 2 or 3. Organizations that plan to keep the model entirely within their own infrastructure and workflows can operate under substantially more permissive terms than those building customer-facing products on top of it.

The community reaction to the release split along predictable lines. Developers praised the breadth of what Moonshot published alongside the weights: attention kernels, MoE communication libraries, agent tooling, and the full technical report. The speed of ecosystem support from inference engines and cloud providers reinforced the perception that this is a substantive release, not a token gesture. Researchers including Nathan Lambert, previously co-leader of Ai2's Olmo model family, pointed out that the license does not fit the usual open-source categories. Lambert described it as "inspired by MIT but distinctly non-commercial," a characterization the license structure supports.

Moonshot is not the first frontier AI developer to adopt this approach. Meta's Llama family uses a community license with a 700 million monthly user threshold for commercial agreements. Other developers have shipped weights under bespoke terms governing redistribution, commercial use, and attribution. Kimi K3 follows that pattern with a different mechanism: rather than restricting redistribution broadly, Moonshot ties commercial rights to company scale and deployment type. The result is a license that feels open for most users but contains obligations that surface only at specific revenue levels or user thresholds.

For engineering and legal teams evaluating Kimi K3, the deployment type determines the compliance burden. Internal-only deployments appear to avoid the commercial licensing provisions entirely. Products that offer API access to third parties must be assessed against the revenue threshold and affiliate scope. Products expecting large-scale consumer deployments must account for the attribution requirement as a product design constraint, not just a legal footnote. The license does not require disclosure of the technical architecture, training methodology, or evaluation results. It requires disclosure of the model name at scale, which is a narrower but still non-trivial obligation.

The release illustrates a distinction that is becoming structurally important across the AI industry. "Open weights" and "open source" are converging terms but do not mean the same thing. Downloading and modifying a frontier model's weights is now straightforward for many organizations. Understanding the legal conditions attached to deploying that model commercially is a separate exercise that requires reading the license, assessing organizational structure, and determining how growth trajectories affect obligations over time. The technical benchmarks and the legal terms are now equally part of the procurement question.