Quantization guides, including Local AI Zone's compression master guide, tell you how to shrink a model. This page answers the question that comes right after: what does the license of the shrunken model let you do with it? We joined our license records to our model records for the same 200 open models. Forty-nine of the 200 publish an official quantization. Here is what their licenses say.
The 49, by format
An official quantization is one the model's own publisher ships, not a community conversion. Formats overlap: a model can publish both a GGUF and an FP8 build.
| Format family | Models publishing it |
|---|---|
| GGUF | 17 |
| FP8 | 17 |
| FP4 variants (incl. NVFP4 and MXFP4) | 13 |
| AWQ | 3 |
| No official quantization published | 151 |
So three quarters of the models in our records leave quantization to the community, and community conversions inherit the base model's license either way. The license question is the same question; the official 49 just make it easy to check.
The format split also dates the field. GGUF is the local-inference standard and the oldest of these families; the FP4 variants are the newest, published by vendors for their own accelerators. A 2024 model that publishes an official quantization usually means GGUF. A 2026 flagship usually means FP8 or FP4, with GGUF left to community converters.
What their licenses allow
Of the 49, 47 allow commercial use and 16 carry extra restriction text. The license split is more permissive than the full 200: Apache 2.0 covers 24 and MIT 11, so 35 of the 49 (71%) sit on the two classic permissive licenses, against 58% across all 200 models.
The two exceptions matter more than the pattern. MiniMax M2.7 ships an official quantization under a license our records classify as non-commercial, and Qwen 2.5 3B Instruct's quantized builds sit under the Qwen license agreement, which our records also flag as barring commercial use. A small, fast, officially quantized model is exactly what an edge product wants, and these two cannot be shipped in one.
The check to run
Before a quantized model goes into a product, read its license record the same way you would the base model's: commercial flag first, restriction text second. Our ShipSafe Filter runs that check against the records behind this page, and our restrictions census shows what the extra clauses usually say.
Sources: Local AI Zone, AI Model Quantization 2026; this site's license and model records for 200 open models (collected September 2025, per-model sources on file). License summaries are orientation, not legal advice.