Anansi-35B-A3B
Anansi-35B-A3B balances good instruct following, narrative reasoning, and sweet prose. Specifically for my 8GB potato, because I can't run the ~30B dense models, but maybe it can be good for your potato too?
This is the repo for Q8 prose/style head GGUFs. The Q8-Prose versions use the same Anansi LM output head — the DARE/TIES merge head interpolated with Melody — but restored at Q8_0 fidelity. The idea is to preserve more of the prose/style effect by avoiding some of the precision loss that comes from quantizing the head along with the rest of the model.
There are also Q8-DarkScarlett versions using the same interpolated-head approach with ReadyArt's Dark Scarlett instead of Melody, giving Anansi an alternate prose/style variant while preserving its underlying reasoning and instruction behaviour. Big thanks to ReadyArt for the donor model.
The Q8-Prose versions are the standard Anansi model with a higher-fidelity prose head. The Dark Scarlett versions are alternate LM-head variants of the same model.
The full Safetensors base model, along with more detail on the merge method and other characteristics, is here.
Recast / additional prose pass
With Anansi in general, and especially with these Q8 prose-head versions if you find the higher-fidelity head gives you a useful prose bump, I recommend at least trying it with the Recast post-processing extension in SillyTavern.
I've found this combination unusually reliable for doing a single additional prose pass. Give the model instructions describing the kind of prose you want (or don't want), tell it to keep its reasoning concise and only output the revision, and keep the original reply reasonably short. I use a fairly high maximum token limit (16k), with my system/character instructions targeting roughly four or five paragraphs.
The model will often still include some reasoning or extra drafts in the result. I just delete those and keep the final revision. It's much easier than editing the prose myself, and in my testing it has been very consistent.
This was actually part of what I was aiming for with Anansi: strong enough instruction following to handle non-continuation tasks like post-processing reliably, while still having the stronger prose and narrative focus I liked in WorldSim. The extra pass can push the prose a fair bit further, although it does mean giving the model enough output space to think and rewrite without getting cut off.
The same general approach may also work with other post-processing workflows or environments that can give the model a revision task rather than asking it to continue the story.
