MichaelAnthony/gemma4-e2b-Snowfox-hf
MichaelAnthony/gemma4-e2b-Snowfox-hf is a 5.1 billion parameter language-only LoRA merge built on Google's Gemma 4 E2B instruction QAT-derived model, developed by Michael Anthony Falabella. This canonical BF16 Transformers checkpoint retains the frozen image and audio towers of the base model, focusing its fine-tuning exclusively on the language backbone. With a context length of 32768 tokens, it serves as the source for various quantized distributions and is optimized for full-precision Transformers inference in language-based applications.
Loading preview...
Overview
MichaelAnthony/gemma4-e2b-Snowfox-hf is the canonical BF16 Transformers checkpoint for SnowFox, a language-only LoRA merge developed by Michael Anthony Falabella. It is built upon Google's Gemma 4 E2B instruction QAT-derived model, specifically the q4_0-unquantized checkpoint. This model serves as the primary source for all SnowFox distributions, including MLX FP16, 4-bit, 6-bit, and GGUF formats.
Key Characteristics
- Architecture: LoRA fine-tune applied only to the language backbone of the Gemma 4 E2B base model.
- Multimodal Towers: The image and audio towers from the base model are frozen and unchanged, preserving their original clipping parameters.
- Parameter Count: Approximately 5.1 billion total parameters, with an effective 2.3 billion parameters.
- Weights Format: BF16, making it suitable for full-precision Transformers inference.
- Context Length: Supports a context length of 32768 tokens.
Use Cases
- Transformers Inference: Ideal for developers requiring full-precision Transformers inference for language-based tasks.
- Derivative Exports: Serves as the foundational source for creating various quantized and optimized model packages for different deployment environments (e.g., MLX, GGUF).