MichaelAnthony/gemma4-e2b-Snowfox-MLX

Gemma 4 E2B SnowFox MLX FP16

This repository contains exactly one MLX variant: the unquantized FP16 SnowFox model. It is a genuine MLX safetensors package, not a GGUF file or a renamed Hugging Face BF16 checkpoint. Four safetensors files make up one model; the shard split is only for reliable large-file download.

SnowFox is a language-only LoRA merge based on Google's Gemma 4 E2B instruction QAT-derived checkpoint. The image and audio towers were frozen during fine-tuning and are retained here, together with the processor and tokenizer needed by MLX-VLM.

Exact lineage

• Base: google/gemma-4-E2B-it-qat-q4_0-unquantized
• Pinned base revision: 6befbaca7398925921802abd1f277b495b78b738
• Canonical merged BF16 source SHA-256: b8fac0ad2cafcb0e7fe29ca6c1deda1389c645751599fe716d4b6f6c0387a2d5
• Conversion: structurally converted to the MLX-VLM v0.6.13 Gemma 4 tensor contract, then cast from BF16 to FP16 for storage.
• Claim boundary: QAT-derived from the base; SnowFox was not trained in FP16 and the post-LoRA weights were not newly QAT-calibrated.

Package contents

• model-00001-of-00004.safetensors through model-00004-of-00004.safetensors: the one FP16 MLX model.
• model.safetensors.index.json: complete shard map.
• config.json, generation_config.json, processor_config.json, tokenizer files, and chat_template.jinja: Gemma 4 E2B multimodal support files.
• mlx_export_manifest.json: source/output provenance and artifact hashes.

Verification performed

The Windows conversion host does not have a compatible MLX runtime, but the stored model conversion was exhaustively verified before upload:

• 1,951 source tensors mapped to 1,951 MLX tensors with no missing or extra keys.
• All 5,104,298,467 stored values were checked after conversion.
• Every output tensor is finite FP16, has exact BF16-to-FP16 values, and its safetensors shard declares format=mlx.
• The largest absolute stored weight is 900.0, below FP16's finite limit.
• The full image/audio/projector tensor set is present; Gemma 4 audio convolution weights use the MLX-VLM axis layout.

Apple-Silicon MLX-VLM inference has not been run from this Windows/AMD release host. Treat this as structurally validated conversion data pending a real Apple-Silicon text, image, and audio generation smoke test; do not interpret the SnowFox training validation scores as fresh MLX runtime results.

Run on Apple Silicon

Use full MLX-VLM, not text-only MLX-LM, because Gemma 4 E2B includes image and audio components:

python -m pip install "mlx-vlm==0.6.13"

python -m mlx_vlm.generate \
  --model MichaelAnthony/gemma4-e2b-Snowfox-MLX \
  --max-tokens 128 \
  --temperature 0.0 \
  --prompt "Explain what SnowFox is in one sentence."

For image prompting, add --image /path/to/image.png to the generation command. Use current MLX-VLM documentation for image, audio, video, and chat-template options.

Quantized variants

Standard MLX-VLM affine quantizations of SnowFox are published as separate repositories and are loadable directly by mlx_vlm.generate:

Variant	Quantization	Size	Notes
gemma4-e2b-Snowfox-MLX-4bit	4-bit affine, group 64	~3.55 GB	GGUF Q4_K_M analogue
gemma4-e2b-Snowfox-MLX-6bit	6-bit affine, group 64	~4.71 GB	GGUF Q6_K analogue

These quantize the language backbone (including the large per-layer embeddings) to 4-bit/6-bit affine while keeping the vision and audio towers dense in FP16, so they are smaller than a standard Linear-only quantization.

The earlier oMLX oQ ("oQ4/oQ6/oQ8") build-to-order plan was never published; use the standard 4-bit/6-bit packages above instead.

License

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