CatAreeya-12B-it in GGUF format, built on Google's Gemma 4 12B IT QAT (Quantization-Aware Trained) checkpoint. Because the base model was trained to tolerate 4-bit quantization, the Q4_0 weights are near-lossless compared to post-training quantization of a standard checkpoint.

The model is multimodal (text + vision) and ships with a Multi-Token Prediction (MTP) drafter for accelerated speculative decoding in llama.cpp.

Files

File	Size	Purpose
`catareeya-4-12B-it-qat-UD-Q4_K_XL.gguf`	6.3 GB	Main model — QAT Q4_0 weights, F32 norms
`mmproj-F16.gguf`	167 MB	Vision projector (required for image input)
`mtp-gemma-4-12B-it-Q4_0.gguf`	242 MB	MTP drafter for speculative decoding

Run with llama.cpp

bash

llama-server \
  --model catareeya-4-12B-it-qat-UD-Q4_K_XL.gguf \
  --host 0.0.0.0 --port 8080 \
  --mmproj mmproj-F16.gguf \
  --model-draft mtp-gemma-4-12B-it-Q4_0.gguf \
  --temp 1.0 \
  --top-p 0.95 \
  --top-k 64 \
  --spec-type draft-mtp \
  --alias catareeya \
  -ctk q8_0 \
  -ctv q8_0 \
  -fa on -cb \
  -np 4

Flag notes

--mmproj — loads the vision encoder; omit it for text-only use.
--model-draft + --spec-type draft-mtp — enables Gemma 4 Multi-Token Prediction speculative decoding for significantly faster generation.
-ctk q8_0 -ctv q8_0 — quantizes the KV cache to Q8_0 to halve context memory.
-fa on — Flash Attention.
-cb — continuous batching.
-np 4 — 4 parallel inference slots.

Download and run directly from the Hub

A recent llama.cpp can pull the files straight from Hugging Face:

llama-server -hf tokenine/catareeya-4-12b-qat

Sampling defaults

Parameter	Value
Temperature	1.0
Top-p	0.95
Top-k	64

Credits

Base model: google/gemma-4-12B-it-qat-q4_0-unquantized
GGUF conversion recipe: unsloth/gemma-4-12B-it-qat-GGUF
License: Gemma Terms of Use