[Qwen] Improve model configuration options for instruct behavior

Hey all, I have been trying to get Qwen 3.5 (Unsloth), under linux, to behave like an instruct model. I have tried a number of permutations, different Qwen 3.5 models, different quants, parameter after parameter. I even tried someone's Jinja template, which supposedly restored the /no\_thinking argument. Nothing has worked. Certainly, Qwen 3 instruct models work, but I am working with limited hardware. Here are the models I have tried: * Qwen3.5-27B-UD-Q8\_K\_XL.gguf * Qwen3.5-27B-Q5\_K\_M.gguf * Qwen3.5-9B-UD-Q3\_K\_XL.gguf * Qwen3.5-122B-A10BQ4\_K\_M-00001-of-00003.gguf * Qwen3.5-35B-A3B-Q4\_K\_M.gguf * Qwen3.5-35B-A3B-UD-Q5\_K\_XL.gguf My usual parameters are: `llama-server --no-mmap -ngl 99 -fa 1 --fit on --seed 3407 --temp 0.7 --top-p 0.8 --min-p 0.0 --top-k 20 --jinja --presence_penalty 1.5 --repeat-penalty 1` And I have tried the following options to disable thinking: * \--chat-template-kwargs '{"enable\_thinking":false}' * \--reasoning-budget 0 * \--chat-**template**\-file Qwen\\ 3.5\\ Jinja\\ template\\ –\\ Restores\\ no\_thinking\\ behavior.txt Any idea what is going on? What am I missing? Thanks!