A text-only derivative of google/gemma-3-12b-it, with the language model extracted from the multimodal checkpoint. The multimodal wrapper (vision tower + multi-modal projector) is removed and the model is reconstructed as a pure Gemma3ForCausalLM (model_type: gemma3_text).

Why

The original multimodal gemma-3 applies bidirectional attention over image tokens (mm-prefix), which prevents it from loading on vLLM's FlashInfer backend. The text-only form clears the mm-prefix flag, so it runs on the FlashInfer / KV-compaction pipeline.

How it was extracted

Keep only language_model.* weights, strip the prefix to model.*
Drop vision_tower.* and multi_modal_projector.*
config = the original text_config (48 layers, sliding_window=1024, 5:1 hybrid)
tokenizer and other files are unchanged from the original

Weight values are identical to the original (lossless for the language model). Image inputs are not supported.

License

This is a Gemma derivative. Use is governed by the Gemma Terms of Use and the Prohibited Use Policy. "Gemma" is a trademark of Google LLC.

PsalMonster/gemma-3-12b-it-textonly

gemma-3-12b-it (text-only)

Why

How it was extracted

License

PsalMonster/
gemma-3-12b-it-textonly