Qwen3.8 suddenly became Qwen3.6?!
After applying this chat template to Qwen3.8 it suddenly became generating code almost the same as Qwen3.6 does. Try it yourself. Here are the steps:
1. Using llama.cpp server start a server without this chat template. Here are my parameters for 2x RTX 5060 TI:
llama-server -m models/esatapedico/Qwen3.8-27B-NVFP4-MTP-GGUF/Qwen3.8-27B-NVFP4-MTP-LOW.gguf --mmproj models/esatapedico/Qwen3.8-27B-NVFP4-MTP-GGUF/mmproj-BF16.gguf -mmdev CUDA1 --jinja -lm mlock -t 6 -ngl 999 -c 262144 -cram -1 -kvu --temp 0.6 --top-k 20 --top-p 0.95 --min-p 0.0 -fa on -np 2 --presence-penalty 0.0 --repeat-penalty 1.00 --image-min-tokens 1024 -ctk q8_0 -ctv q8_0 --spec-type ngram-mod,draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0.5 -sm tensor -fit off -b 2048 -ub 256 -dev CUDA0,CUDA1 --main-gpu 0 --reasoning-preserve
2. Open llama.cpp webUi, start new chat, set reasoning to medium, enter prompt: "make beautiful snake game in single HTML file."
This is what i get. The title is some fancy name for a snake game and the look is much different what Qwen3.6 generated. Total tokens generated is about 15k.
3. Now start new llama.cpp server with this template (v22.3). Here are the same settings again, but with added recommended template parameters:
llama-server -m models/esatapedico/Qwen3.8-27B-NVFP4-MTP-GGUF/Qwen3.8-27B-NVFP4-MTP-LOW.gguf --mmproj models/esatapedico/Qwen3.8-27B-NVFP4-MTP-GGUF/mmproj-BF16.gguf -mmdev CUDA1 --jinja --chat-template-file models/esatapedico/Qwen3.8-27B-NVFP4-MTP-GGUF/chat_template.jinja --reasoning-format deepseek -lm mlock -t 6 -ngl 999 -c 262144 -cram -1 -kvu --temp 0.6 --top-k 20 --top-p 0.95 --min-p 0.0 -fa on -np 2 --presence-penalty 0.0 --repeat-penalty 1.00 --image-min-tokens 1024 -ctk q8_0 -ctv q8_0 --spec-type ngram-mod,draft-mtp --spec-draft-n-max 4 --spec-draft-n-min 0.5 -sm tensor -fit off -b 2048 -ub 256 -dev CUDA0,CUDA1 --main-gpu 0 --reasoning-preserve
4. Open llama.cpp webUi, start new chat, set reasoning to medium again, enter prompt: "make beautiful snake game in single HTML file."
This looks like Qwen3.6 output! Even the title is the same like from Qwen 3.6 - "NEON SNAKE"
I had no idea that this template would have such a big impact on how the Qwen3.8 model reasons and generates output. Total tokens generated is about 6k.
Could you say that is more token efficient because it produced the output with 1/3 of the tokens. Maybe you give a vague input without the template it has more freedom/creativity with the template is just doing what you are asking. Just guessing.
By the way his template has thinking level medium by default, without the template would be xhigh
Could you say that is more token efficient because it produced the output with 1/3 of the tokens. Maybe you give a vague input without the template it has more freedom/creativity with the template is just doing what you are asking. Just guessing.
By the way his template has thinking level medium by default, without the template would be xhigh
Both runs where on reasoning MEDIUM. I tried it few times. With the template the game is always in blue colors and the title is "Neon Snake", just like from Qwen3.6 😃
And you are right, token efficiency is much better, but it's so strange, that without the template colors are green, and with it they are blue - on the same reasoning settings.
The template reduces thinking, Qwen does better the more it thinks. So when "designing" things from scratch you should probably try xhigh or skip the template.