The words "custom license" should change how you read a model page. Moonshot AI's Kimi K2 earned its reputation on strong agentic performance, and the licensing question followed immediately. Kimi K2 ships under a custom license, not a standard permissive one. That sentence should trigger caution, not confidence, because the Kimi K2 license commercial restrictions deserve a real read.
Custom lab licenses usually allow research and commercial use with conditions: attribution, an acceptable use policy, sometimes restrictions on redistribution or on using outputs to train competing models. The exact mix varies, which is precisely why "custom" is a warning label.
The clause founders miss most is the competing-model-training restriction. If your roadmap includes distillation or training on synthetic data, this clause can bite. Check for it explicitly. Any open source LLM licensing guide will flag this as the first thing to grep for in a custom license. I would add one more search: the redistribution terms, because custom licenses love to qualify them.
If you are evaluating Kimi against alternatives, do the license comparison before the benchmark comparison. Benchmarks tell you which model is better. Licenses tell you which one you are allowed to ship. The second question decides whether the first one matters.
My stance on custom licenses generally: they are fine. Just read them. The risk is not that they are bad. It is that people assume Apache 2.0 because the weights are downloadable. Read the file. An AI model commercial rights lookup tool will give you a summary, but for custom licenses the summary is a starting point, not a verdict. AI model license explained properly means going to the source on these.
And document your decision. One paragraph: why this checkpoint, which license, what it permits. When someone asks in six months why you chose it, you will have the answer instead of a vague memory. Diligence loves paper trails.
I compare custom model licenses against standard ones in plain language at localaitools.fyi, with clearance reports for teams that want the fine print handled.