froggeric/Qwen-Fixed-Chat-Templates · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
LLM Prompt Template
Summary
A drop-in chat template for Qwen 3.5, 3.6, and 3.8 models that appends an 11-line terseness prompt to reduce thinking tokens and output padding while preserving accuracy. Claims local Qwen models match or exceed cloud frontier models on SWE-bench-Live.
Highlights
- Eleven-line terseness prompt that cuts filler without sacrificing correctness, making models more intelligent per token
- Claims Qwen3.8-27b (Dirk) out-solves cloud-frontier Opus 5 (high) and Qwen3.6-35B-A3B (Nail) matches Sonnet 5 (high) on SWE-bench-Live
- Supports multiple runtimes: MLX, transformers, GGUF, oMLX, and llama.cpp via a single chat_template.jinja file
- Built on froggeric's v22.1 base which adds reasoning-effort steering (none through xhigh) and inline think control tags
- Portable template decoupled from specific model builds, making it reusable across any compatible Qwen deployment
auto-generated
Context
Audience
Developers and AI practitioners running local LLMs, particularly those using Qwen models for coding and knowledge work who want to optimize token efficiency and reduce inference time
DomainLarge Language Models and Prompt Engineering
FormatOpen source repository with documentation
AccessFree, open source on Hugging Face
Source Overview
This repository provides a drop-in chat template for Qwen 3.5, 3.6, and 3.8 models that appends an 11-line terseness prompt to froggeric's Qwen-Fixed-Chat-Templates v22.1.
The template reduces thinking tokens and output padding while preserving accuracy, and enables thinking retention to avoid loops and improve cache hits across turns. Installation supports MLX, transformers, GGUF, oMLX, and llama.cpp via a single chat_template.jinja file or embedded template rewrite.
Topics
Related
froggeric/Qwen-Fixed-Chat-TemplatesQwen 3.5 / 3.6 / 3.8 modelsSWE-bench-Livellama.cppMLXGGUF quantization
