This repository provides a structurally extracted, text-only iteration of Google's multimodal gemma-4-E4B-it model. Vision and audio encoders have been fully decoupled to minimize VRAM footprint for text-centric workloads.

Note on Zero-Shot Modality Queries: The text parameters retain their original RLHF conditioning. The model will assert multimodal capabilities (e.g., confirming it can interpret images) despite hardware encoders being purged. Overriding this behavior requires explicit bounding via the system prompt.