gemma-4-12B-it -- Pollard
Pollard shrank this model: 23.81 GB (f16) -> 4.64 GB -- 80% smaller, 5.1x down.
The smallest rung here; larger, higher-fidelity rungs are listed below.
format this model's size f16 23.81 GB Q8_0 ~12.62 GB Q6_K ~9.76 GB Q4_K_M ~6.91 GB PollardMix (this repo's IQ2_XXS) 4.64 GB
Pollard builds of google/gemma-4-12B-it made with Pollard Weights -- a ladder of measured-allocation quants (bits placed by per-layer sensitivity, not a uniform crush).
Standard GGUF -- every file here runs in stock llama.cpp / ik_llama.cpp, Ollama, LM Studio.
Model details
| Parameter count | ~11.9B |
| Architecture | gemma4_unified |
| Input support | text, image, audio |
| imatrix | yes -- see calibration |
| Perplexity measured | yes -- table below |
Which file should I choose?
Every rung is the same weights, sized to a different RAM budget by the measured allocation. Pick the largest one that fits your machine with room for context:
- ~12 GB RAM / VRAM ->
Q6_K(9.65 GB). - ~9 GB RAM / VRAM ->
IQ4_XS(6.78 GB). - ~7 GB RAM / VRAM ->
IQ2_XXS(4.64 GB).
Available files (Calib 3.0 held-out (prose/code/math/chat/multilingual), ctx 2048, 60 chunks)
f16 reference PPL 23.041768.
| file | PPL | size | Mean KLD | notes |
|---|---|---|---|---|
gemma-4-12B-it-Pollard-IQ2_XXS.gguf | 56.8803 | 4.64 GB | 2.3671 | smallest -- +147% vs f16 |
gemma-4-12B-it-Pollard-IQ4_XS.gguf | 29.803 | 6.78 GB | 0.7497 | recommended default -- +29% vs f16 |
gemma-4-12B-it-Pollard-Q6_K.gguf | 26.6124 | 9.65 GB | 0.334 | highest fidelity here -- +15% vs f16 |
Download a specific file
pip install -U "huggingface_hub[cli]"
hf download PollardWeights/gemma-4-12B-it-Pollard \
--include "gemma-4-12B-it-Pollard-IQ2_XXS.gguf" --local-dir ./
How to run
These are standard GGUF and run with llama.cpp:
llama-server -hf PollardWeights/gemma-4-12B-it-Pollard:IQ2_XXS
or from a local file:
llama-cli -m gemma-4-12B-it-Pollard-IQ2_XXS.gguf -ngl 99 -p "Explain why the sky is blue."
llama-server -m gemma-4-12B-it-Pollard-IQ2_XXS.gguf -ngl 99 # OpenAI-compatible API + web UI at :8080
They also work in anything built on llama.cpp -- LM Studio, koboldcpp, Jan, ramalama, Ollama (ollama run hf.co/PollardWeights/gemma-4-12B-it-Pollard).
ARM / AVX
llama.cpp repacks weights into an interleaved layout at load time for faster inference on ARM and AVX machines -- no special file needed, online repacking covers these quants. The old Q4_0_4_4/4_8/8_8 variants are not required.
Errata
- Every file here loads in stock llama.cpp -- verified from the tensor types with
pollard-ggufcheck, not assumed from the filenames. - Measured allocation places bits by per-layer sensitivity under a size budget.
- Single machine; replication invited.
Credits & license
- Base model:
google/gemma-4-12B-it - Quantization tooling: llama.cpp (ggml-org)
- Method + tooling: Pollard Weights -- measure first, no claim before a number.
- License:
apache-2.0, inherited from the base model.
Built with Pollard Weights -- frontier models, small hardware, no compromise.