WhiskyAKM/Gemma-4-12B-it-qat-NVFP4-GGUF

Gemma 4 12B IT QAT NVFP4 — GGUF

NVFP4 GGUF conversions derived from WhiskyAKM/Gemma-4-12B-it-qat-GGUF, which was created from the Quantization-Aware Training (QAT) checkpoint of Gemma 4 12B IT (google/gemma-4-12B-it-qat-q4-0).

This repository provides NVFP4 (NVIDIA 4-bit floating-point) and bf16 GGUF files, making the model usable with llama.cpp and other GGUF-compatible inference engines.

Model Overview

Gemma 4 12B IT is a multimodal model built by Google DeepMind that handles text, image, and audio inputs and generates text output. It is designed for efficient on-device and server deployment.

Model Architecture

GGUF Files

A chat_template.jinja file is also provided for use with chat-based inference.

About NVFP4: NVFP4 is NVIDIA's 4-bit floating-point format (E2M1) with microscaled per-block scaling factors. It offers significant memory savings while preserving quality close to bfloat16. NVFP4 inference is accelerated on NVIDIA Blackwell GPUs and supported via llama.cpp on compatible hardware.

Note on mmproj: The mmproj.gguf file contains the vision and audio projectors needed for multimodal (image/audio) inference. It is shared across both quantization variants.

Usage

llama.cpp (CLI)

# Run text-only inference
./llama-cli \
  -m gemma-4-12b-it-qat-nvfp4.gguf \
  -p "Explain quantum computing in simple terms." \
  --temp 1.0 --top-k 64 --top-p 0.95

llama-server (OpenAI-compatible API)

# Text-only
./llama-server \
  -m gemma-4-12b-it-qat-nvfp4.gguf \
  --host 0.0.0.0 --port 8080

# Multimodal (image + audio)
./llama-server \
  -m gemma-4-12b-it-qat-nvfp4.gguf \
  --mmproj mmproj.gguf \
  --host 0.0.0.0 --port 8080

Multimodal (Image / Audio)

For image and audio inputs, use llama-server or llama-cli with the --mmproj flag pointing to mmproj.gguf. Refer to your inference engine's documentation for passing image/audio data alongside text prompts.

Modality order tip: For best results, place image content before text and audio content after text in your prompt.

Generation Parameters

Recommended parameters from the model's generation_config.json:

Thinking Mode

Gemma 4 supports configurable thinking (reasoning) mode:

• —Enable: Include the <|think|> token at the start of the system prompt.
• —Output format: When thinking is enabled, the model outputs internal reasoning followed by the final answer:

  <|channel>thought
  [Internal reasoning]
  <channel|>
  [Final answer]

• —Disable: Omit the <|think|> token.

Many libraries like Transformers and llama.cpp handle the chat template complexities automatically.

Key Features

• —Multimodal: Text, image, and audio understanding
• —Long Context: 256K token context window
• —Function Calling: Native support for structured tool use (agentic workflows)
• —Multilingual: Support for 140+ languages
• —Native System Prompt: Supports the system role for structured conversations

Acknowledgements

• —Original model: google/gemma-4-12B-it
• —QAT checkpoint: google/gemma-4-12B-it-qat-q4-0
• —Base GGUF conversion: WhiskyAKM/Gemma-4-12B-it-qat-GGUF
• —Technical report: Gemma 4 Technical Report (arXiv:2607.02770)

Citation

@misc{gemmateam2026gemma4,
      title={Gemma 4 Technical Report},
      author={Gemma Team},
      year={2026},
      eprint={2607.02770},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.02770},
}

License

Apache License 2.0