The Den Reviewthe briefing for executives who run AI
Wednesday, November 11, 2026

The Denominator

Subscribe
Trending

Model Labs · Open-Weight Models

Moonshot Opens the Weights of a 2.8 Trillion-Parameter Model, With Strings Attached

Kimi K3.1 Pro's license restricts some commercial uses, unlike Apache or MIT. Its API charges $2 in, $10 out.

· 2 min read

A circuit board, chained to a rough stone plinth, stands under a spotlight in a dark loft, as two men in overcoats talk nearby.
Photograph: AI-generated

Any company that sells access to Kimi K3.1 Pro and earns more than $20 million over 12 months, counting its affiliates, must sign a separate agreement with Moonshot AI first, the model’s license says.

That test decides who can build a service on the downloaded weights and who has to negotiate with Moonshot. The license sets no price for that agreement.

Who needs a deal?

The license calls the trigger “Model as a Service”: giving outsiders access to the model’s inference or fine-tuning, through an application programming interface (API) for example, in a way that lets them steer the inputs, settings or training data. A product that embeds the model in specific features, or that only relays requests to models hosted elsewhere, falls outside the definition.

A second condition asks for a label, not a deal. A commercial product built on the model with more than 100 million monthly active users, or more than $20 million in monthly revenue, must show “Kimi K3.1-Pro” prominently in its interface. Both conditions fall away for internal use and for use through Moonshot’s own products or its certified inference partners, a term the license does not define.

Is the model worth the license?

Moonshot’s model card puts Kimi K3.1 Pro at 65.7 on the coding benchmark Terminal-Bench 4.0, under a point behind the 66.4 for Anthropic’s Claude Opus 5.5, and ahead of Opus 5.5 on a second coding benchmark, SWE Marathon, 51.7 to 46.3. Artificial Analysis puts the model at 67.2% on Terminal-Bench 4.0.

Those scores put the next question on the bill: which route a buyer runs the model by.

What does each route cost?

Moonshot’s API costs $2 per million input tokens and $10 per million output tokens, according to pricing guides that quote its rate card. For self-hosting, Nvidia’s page for the model names eight Blackwell GB300 graphics processing units (GPUs) as its test setup and gives no cost for them.

A third route runs through OpenRouter, a marketplace that sends requests to outside hosting companies, 17 of which serve the model. The license names no certified inference partners, so a buyer would have to ask Moonshot whether any of the 17 qualify.

The license’s own test for a deal is whether outsiders steer the model through something the company sells.

NVIDIAMicrosoftOpenAIAnthropicGoogleAWSJPMorganCapital OneAT&TCharterComcastT‑MobileUnitedHealthCVS HealthTravelersWalmartSalesforceDatabricksSnowflakeCoreWeaveStripeAccenture