DoppelReflEx/QWQ-32B-Dawnwhisper-QWQTokenizer

QWQ-32B-Dawnwhisper-QwenTokenizer

This version is make with QWQ Tokenizer. It should be better than Origin version in Reasoning mode, that's all.

Like many people said: Tiny Deepseek R1 at home if you don't have too good specs. 16GB Vram card could run IQ3 variants very well.

After quick test, this merge perform very good result and strong capability in roleplay. Nothing more. Thank you for using my merge model.

After full testing, I can say this model is one of the best models I have created (merged). Dialogues likely vivid and very smart compare to its size. Thank mradermacher's team for quantizing this model, that help me available to do full testing.

The cons are: anything but not English and Chinese, will decrease the performance of this model, you should turn on reasoning mode to have better experience in non-English and non-Chinese language.

GGUF (Thank mradermacher and his team, especially nicoboss)

Static - Imatrix

Setting

Please use ChatML template

Reasoning is not necessary to turn on, but a nice feature to enable if you want to 'boost' your experience when roleplaying and multitasking (but more time consuming :/).

Reasoning token is <thinking> </thinking>. You could search how to enable thinking mode in internet. Note that in silly tarven, you should turn off Always add character's name to prompt in Context Formatting and Include names Never in Instruct Template.

Configuration

The following YAML configuration was used to produce this model:

models:
  - model: trashpanda-org/QwQ-32B-Snowdrop-v0
    parameters:
      density: 0.9
      weight: 1
  - model: ArliAI/QwQ-32B-ArliAI-RpR-v3
    parameters:
      density: 0.8
      weight: 0.8
merge_method: dare_ties
base_model: Qwen/QwQ-32B
parameters:
  normalize: true
  rescale: true
tokenizer_source: Qwen/QwQ-32B
dtype: bfloat16