Thanks but still not good enough unfortunatly in creative workflows.
I’ve already conducted several tests and evaluations with Astra to compare the results of this model with those of the Qwen3.8-27B-Uncensored-GGUF (JonathanColetti) model, GGUF Q4 K_M. JonathanColetti’s version consistently outperforms despite its smaller size. I use it in a ComfyUI workflow with a local Llama server, the model manages several custom nodes and validators to generate story bibles and shot plans from a simple (or detailed) idea, which are then passed to a node that creates prompts for the H3 video model. Unfortunately, the Ultra Heretics version (and the previous ones—I’ve tested them all) still fails to produce correct and creative JSON from start to finish, even when the recommended settings are applied...
Thanks anyway for the effort, no doubt it must take a lot of time.
Hey
You need Q6 or Q8 ; due to reasoning reductions and/or slightly more detailed prompts.
Qwen3.8-27B-Uncensored-GGUF (JonathanColetti) is the full Qwen 3.8, untuned in heretic form.
OR you may want to try Fable Fusion 711 [Qwen 3.6] as this is a far more extensive tune; and very strong in all areas.