[!TIP] KV-cache quantization (upstream, no fork needed): llama.cpp/Ollama cover this natively — -ctk q8_0 -ctv q8_0 (~half KV memory, negligible quality loss) or -ctk q4_0 -ctv q4_0 (~quarter memory, small quality cost). In Ollama: OLLAMA_KV_CACHE_TYPE=q8_0 with OLLAMA_FLASH_ATTENTION=1.

gemma-4-12B — TurboQuant GGUF IQ4_XS

google/gemma-4-12B @ 023679ed352de9bb66cc873c9009ce3482585c08 quantized pack, published as majentik/gemma-4-12B-TurboQuant-GGUF-IQ4_XS.

Method

llama.cpp IQ4_XS quantization, text-only (vision/audio not included).

Release line

Released under the TurboQuant line. RotorQuant and TurboQuant are this project's release labels for this pack, not distinct quantization algorithms — both brand repos for a given tier carry byte-identical weights, produced once and published under two names. No brand-specific speedup is claimed or measured for either label.

Modality

TEXT-ONLY. This is a llama.cpp GGUF conversion of the text tower only — the vision and audio towers are NOT included in this pack. Do not expect image or audio understanding from this repo; use one of the MLX packs for multimodal input.

License

Governed by the Gemma Terms of Use. See the upstream repo for the full license text.