Three sizes of Gemma 3, all Q4_K_M from the ggml-org conversions, all verified ungated single-file GGUFs:
| Catalog id | Repo | Size |
|---|---|---|
gemma-3-1b-it | ggml-org/gemma-3-1b-it-GGUF | 0.81 GB |
gemma-3-4b-it | ggml-org/gemma-3-4b-it-GGUF | 2.5 GB |
gemma-3-12b-it | ggml-org/gemma-3-12b-it-GGUF | 7.3 GB |
One honest flag: llama.cpp has no native tool-call parser for the Gemma family yet, so these entries ship with tools: no. Requests that demand tool calls will get a clean 400 unsupported rather than silently broken output — that is the catalog's contract: the flag tells the truth.
The 12B at 7.3 GB is the largest entry in the catalog today; give it 10 GB of memory headroom. The 1B is a fine default for feature prototyping where download time matters more than depth. The 4B has since gained a multimodal projector in the catalog and takes image input — see the v0.4.0 notes.