Gemma 4 31B sparks writer-model debate

My take: the base checkpoint is probably good enough to surprise people, but the best creative-writing experience will come from finetunes that sand off its more clinical voice.

• –Google positions Gemma 4 31B as the quality-first dense model; the 26B A4B is the practical efficiency pick, which matters a lot once you count KV cache and long context.
• –Official materials emphasize reasoning, long context, and agentic workflows more than literary style, so raw prose quality is only part of the story.
• –Community finetunes like ConicCat/Gemma4-Garnet-31B explicitly target prose, roleplay, and lower-slop outputs, making them more relevant if the goal is fiction or stylized writing.
• –On 16GB-class hardware, a quantized 31B can be made to work, but the experience will be tighter and less flexible than a smaller model with more headroom.
• –If the question is “best writing per GB,” Gemma 4 is a serious contender, but not an automatic win over smaller, more style-focused local models.