Q8,FP16?
It'd be great to see Q8&FP16 in a V2? Also: Image capability? does image capability break flash attention?
It'd be great to see Q8&FP16 in a V2?
You can find the Q8 model under https://huggingface.co/mradermacher/Fallen-Gemma3-27B-v1-GGUF. Wighted/imatrix quants do NOT exist for Q8. FP16 we only provide for tiny models as even Q8 is already pointless in terms of quality from any practical use case. If you don't beleve me take a look at the quant quality measurements I posted under https://huggingface.co/mradermacher/BabyHercules-4x150M-GGUF/discussions/2 (enable hidden messages)
Also: Image capability?
Regarding image capability the original model is Gemma3ForCausalLM and so has no image capabilities. For it to have image generation capabilities it would need to be of type Gemma3ForConditionalGeneration.
image capability break flash attention
No it does not as far I'm aware.