Repository navigation
#enhancement using jinja template from "peculiar-ragdoll/Qwen-Sharp-Chat-Templates" Wow!!! #99
Description
Im using Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS and I replaced both jinja templates with the one from peculiar-ragdoll (This is a drop-in fix for any Qwen3.5, 3.6, or 3.8 model, optimizing the models for knowledge work and coding) and wow its a night and day difference to the base jinja template. Try it for yourself.
peculiar-ragdoll/Qwen-Sharp-Chat-Templates
Location of the base jinja templates that I replaced:
Strata-data/packs/iq3_xxs/tokenizer/chat_template.jinja
Strata/serve/chat_template.jinja
Hope that helps!
Reacted by ukrolelo
Activity
Niko1221 commentedon Sep 29, 2026
Thanks for sharing it @technerd-official. Strata keeps the model's own chat template, so answers match what the model was trained with; I won't ship a third-party template. Part of the looping on images had a real cause here (image requests ignored the sampling settings - fixed for 0.1.23, see #75). You can still use your template with a client that renders it itself. Closing.