Every lab writes its own licence

Apache 2.0, MIT, community licences with tripwires, revenue share, and now a licence named after the model it governs. Five legal regimes, one phrase, and the phrase has stopped meaning anything.

Kimi K3 — 2.8 trillion parameters, billed as the largest open-source model — shipped under the "Kimi K3 License", three weeks after Alibaba attached a revenue share to its own frontier weights.

Count them

Permissive (Apache 2.0, MIT). Community licences with acceptable-use appendices and user-count thresholds. Revenue-share grants. Per-model bespoke instruments. All of it published under the words "open weights", and only the first category lets an engineer ship without a legal review.

A licence with no precedent is a project

A bespoke instrument named after one model has no case law, no community interpretation, and no prior art. Every organisation evaluating it does the analysis from scratch, and the answer does not generalise to the next release.

Openness of weights and openness of terms have fully decoupled. Only the second determines whether you can ship.

Why labs are doing it

Because permissive licensing gives away the commercial position entirely, and the labs that adopted it did so when they were behind and needed adoption more than revenue. A lab at 2.8 trillion parameters with genuine developer trust is in a different negotiating position, and bespoke terms are how that position gets expressed.

The same logic is visible in pricing: DeepSeek is raising prices while Western labs cut. Cheap and permissive was a strategy for entering the market. Neither is a strategy for staying in it.

Which is fine and rational — but the sector should stop using one word for five contracts.

Layer3 Labs — Best Chinese AI Models 2026: Kimi K3, DeepSeek, Qwen → · OpenCurious — Top 33 Open-Source AI Models (2026) → · LLM Stats — AI Updates Today (August 2026) →