llama.cpp quantizations of andyoneal/Gemma-4-E4B-Nightcap, a roleplay / creative-writing task-arithmetic merge of Gemma 4 E4B. See the main repo for the full card — what's in it, how it behaves, and how the recipe was chosen.

Text-only (the conversion drops the audio/vision towers); chat template embedded.

file	size	notes
`Gemma-4-E4B-Nightcap-Q8_0.gguf`	7.5 GB	Recommended. The tier the model was blind-tested at; one parent's voice audibly degrades below Q8.
`Gemma-4-E4B-Nightcap-Q6_K.gguf`	5.8 GB	Desktop compromise tier.
`Gemma-4-E4B-Nightcap-Q4_K_M.gguf`	5.0 GB	Low-RAM tier.
`Gemma-4-E4B-Nightcap-Q4_0.gguf`	4.8 GB	Phone tier — llama.cpp repacks Q4_0 into ARM-optimized layouts at load, so it's the fast pick in mobile apps.

Samplers: temp 1.0, min-p 0.02, no repetition penalties (they do more harm than good on this family). Temp 0.8 / top-p 0.95 also works fine.

andyoneal/Gemma-4-E4B-Nightcap-GGUF

Gemma 4 E4B Nightcap — GGUF

andyoneal/
Gemma-4-E4B-Nightcap-GGUF