Gemma 4 E4B IT QAT NVFP4 — GGUF

NVFP4 (NVIDIA Floating Point 4-bit) GGUF quantization of google/gemma-4-E4B-it-qat-q4_0-unquantized, created from the Quantization-Aware Training (QAT) checkpoint of Gemma 4 E4B IT.

Model Overview

Gemma 4 E4B 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 deployment on laptops and mobile devices.

The "E" in E4B stands for "effective" parameters — the model uses Per-Layer Embeddings (PLE) to maximize parameter efficiency, giving each decoder layer its own small embedding for every token. This keeps the effective parameter count much smaller than the total.

This repository contains an NVFP4 GGUF conversion of the QAT-optimized checkpoint, making it usable with llama.cpp and other GGUF-compatible inference engines that support the NVFP4 data type.

Model Architecture

GGUF Files

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

Note on NVFP4: NVFP4 uses 4-bit floating-point representation (1 sign + 3 mantissa + 2 exponent bits on an E2M1 basis with microscaling) designed for NVIDIA Blackwell and later GPUs. It offers significant memory savings while maintaining good quality thanks to the QAT optimization.

Note on mmproj.gguf: This file contains the vision and audio encoders/projection layers required for multimodal inference. It is not needed for text-only use.

Usage

llama.cpp (CLI)

# Run text inference
./llama-cli \
  -m gemma-4-E4B-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-E4B-it-qat-nvfp4.gguf \
  --host 0.0.0.0 --port 8080

# With multimodal (image + audio) support
./llama-server \
  -m gemma-4-E4B-it-qat-nvfp4.gguf \
  --mmproj mmproj.gguf \
  --host 0.0.0.0 --port 8080

Multimodal (Image / Audio)

For image and audio inputs, pass the --mmproj mmproj.gguf flag to load the multimodal projector. 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. For the E4B variant, thinking is fully off when disabled (no empty thought block is generated).

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

Key Features

• Multimodal: Text, image, and audio understanding
• Long Context: 128K token context window
• Function Calling: Native support for structured tool use (agentic workflows)
• Multilingual: Support for 140+ languages
• On-Device Optimized: Designed for efficient local execution on laptops and mobile devices
• Native System Prompt: Supports the system role for structured conversations

Acknowledgements

• Original model: google/gemma-4-E4B-it
• QAT checkpoint: google/gemma-4-E4B-it-qat-q4_0-unquantized
• 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