---
base_model: google/gemma-4-E2B-it-qat-q4_0-unquantized
library_name: transformers
pipeline_tag: image-text-to-text
license: apache-2.0
tags:
- gemma4
- snowfox
- qat-derived
- safetensors
- transformers
- multimodal
- text-generation
---

# Gemma 4 E2B SnowFox (canonical BF16 source)

This is the **canonical merged BF16 Transformers checkpoint** of SnowFox — a
language-only LoRA merge built on Google's Gemma 4 E2B instruction QAT-derived
model. Every SnowFox distribution (MLX FP16, MLX 4-bit, MLX 6-bit, GGUF) is
derived from this repository, so this is the package to use for full-precision
Transformers inference or as the source for your own exports.

SnowFox is trained by **Michael Anthony Falabella**.

## What SnowFox is

SnowFox is a language-only LoRA merge: the **image and audio towers were frozen**
during fine-tuning and are retained unchanged from the base model. Only the
language backbone received the SnowFox LoRA adaptation. The base is Google's
QAT-derived `q4_0-unquantized` checkpoint, which carries clipping parameters on
the multimodal towers that are preserved here.

## Model size

| Property | Value |
|---|---|
| Total parameters | ~5.1B (with per-layer embeddings) |
| Effective parameters | ~2.3B |
| Weights format | BF16 |
| Checkpoint size | ~10.2 GB (`model.safetensors`) |

> Note: Hugging Face's model page may report a smaller "params" figure for the
> quantized MLX derivatives of this model. That is a display artifact — those
> repos store weights as *packed* `uint32` words (8× 4-bit / 5× 6-bit values per
> word) and HF counts each packed word as one parameter. The true count is
> unchanged (~5.1B total / ~2.3B effective).

## Exact lineage

- Base: [`google/gemma-4-E2B-it-qat-q4_0-unquantized`](https://huggingface.co/google/gemma-4-E2B-it-qat-q4_0-unquantized)
- Pinned base revision: `6befbaca7398925921802abd1f277b495b78b738`
- Method: LoRA fine-tune (language-only), merged into the base model
- Claim boundary: QAT-derived from the base; SnowFox's post-LoRA weights were
  not newly QAT-calibrated.

## Quick start

```python
from transformers import AutoModelForCausalLM, AutoProcessor

model_id = "MichaelAnthony/gemma4-e2b-Snowfox-hf"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)
```

## Derivative packages

| Package | Format | Notes |
|---|---|---|
| [`gemma4-e2b-Snowfox-MLX`](https://huggingface.co/MichaelAnthony/gemma4-e2b-Snowfox-MLX) | MLX FP16 | `mlx-vlm` ready |
| [`gemma4-e2b-Snowfox-MLX-4bit`](https://huggingface.co/MichaelAnthony/gemma4-e2b-Snowfox-MLX-4bit) | MLX 4-bit affine | ~3.55 GB |
| [`gemma4-e2b-Snowfox-MLX-6bit`](https://huggingface.co/MichaelAnthony/gemma4-e2b-Snowfox-MLX-6bit) | MLX 6-bit affine | ~4.71 GB |
| [`gemma4-e2b-Snowfox-GGUF`](https://huggingface.co/MichaelAnthony/gemma4-e2b-Snowfox-GGUF) | GGUF | llama.cpp / Ollama |

## License

Gemma 4 is Apache-2.0. This derivative package uses the Apache-2.0 license
declared by the pinned base model. See [`LICENSE`](LICENSE) and
[`NOTICE.md`](NOTICE.md) for the lineage and modification notice.