gemma-4-12B-it · GGUF

A mid-size Gemma 4 model with image input, configurable thinking and a 256K context.

GGUF builds of gemma-4-12B-it by Google for llama.cpp and SoAI, calibrated on real chat, code and tool calls.

---

At a glance

Choosing a file

[!TIP] Q4_K_M suits most machines. It is about a third of the size of the original 16-bit weights and stays close to them in quality. If Q8_0 fits in your memory, it keeps output closest to the original. SoAI does not publish quants below 4 bits because their quality loss is rarely worth the saved space.

The model needs memory for the file and for the context, and the context part grows with the context length you set. llama.cpp can run the model on the GPU, on the CPU, or with its layers split between them.

Images need the mmproj-gemma-4-12B-it-bf16.gguf add-on as well, loaded with --mmproj. llama-server -hf downloads it automatically.

The add-on includes the model's audio encoder, but llama.cpp up to v0.5.0 does not transcribe audio reliably with it yet. Image input works, and Gemma 4 E2B and E4B handle speech.

Quick start

With llama.cpp, this command starts a local server with a built-in chat UI and an OpenAI-compatible API:

llama-server -hf SoAIHQ/gemma-4-12B-it-GGUF:Q4_K_M

Open http://localhost:8080 to chat, or send requests to the API:

curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
  "messages": [{"role": "user", "content": "Explain GGUF in one sentence."}],
  "chat_template_kwargs": {"enable_thinking": true}
}'

Download the files to a folder first

hf download SoAIHQ/gemma-4-12B-it-GGUF --include "gemma-4-12B-it-Q4_K_M.gguf" "mmproj-*" --local-dir gemma-4-12B-it-GGUF
llama-server -m gemma-4-12B-it-GGUF/gemma-4-12B-it-Q4_K_M.gguf \
  --mmproj gemma-4-12B-it-GGUF/mmproj-gemma-4-12B-it-bf16.gguf

Recommended settings

Google recommends these values in the original model card. Google uses the same sampling settings for every task, with thinking on or off.

In llama-server, you set reasoning per request with "chat_template_kwargs": {"enable_thinking": true}.

How these files are made

Q4_K_M is quantized with an importance matrix (imatrix), which tells llama.cpp which weights affect the output most so it can store them with more precision. SoAI computes the imatrix from its own calibration corpus instead of generic web text, and formats every conversation in it with this model's chat template, including its native tool-call syntax. The model is therefore measured on the token patterns it sees in use: multi-turn chat in 21 languages, code edits in 22 programming languages, tool calls and their results, step-by-step math, and web prose.

Q8_0 is quantized without the imatrix, because llama.cpp's Q8_0 format does not use one.

Before any quantization, the build checks that the chat, thinking and tool-call markers are stored as special tokens and that the embedded chat template matches the original. A conversion that imports those markers as plain text breaks turns, thinking and tool calls without any error message, so the build stops instead of publishing the file.

The imatrix file is published in this repository, and the provenance table below pins the original revision and the llama.cpp commit, so you can rebuild these files yourself.

Calibration corpus sources All files and SHA-256 checksums

The full checksums are in SHA256SUMS.

Provenance

License and credits

These files are distributed under the original model's license (apache-2.0), and its full terms continue to apply. The license and the original model card also describe the model's capabilities, training data, benchmarks and limitations.

Google designed and trained the model, and SoAI only repackages the published weights. The conversion uses llama.cpp by Georgi Gerganov and the ggml contributors.

Run it with SoAI

Your models, your hardware, one software suite. Download SoAI for Windows, macOS, Linux or Android.

_{Packaged by SoAI.}