Qwen-3.5-2b-roleplay-tuned-v2-ERP-tolerant

My newest complex roleplay finetune of Qwen 3.5 2b, trained on more data with improved methodology, aimed at enhancing creative writing and changing the overall writing style.

Overview:

This model has:

• Improved creative writing

• Improved punctuation

• Improved tolerance to explicit themes/Reduced refusals

• Possible instability

Note:

No quants for the model this time, but:

• BF16: Highest quality

• INT8W8A16: Very high quality

• FP8: Very high quality

• NVFP4: High quality

UPD: Added new safetensors weights:

• INT8W8A8: Very good quality

• INT4W4A16: Alright quality, small memory footprint

Theres also LoRA adapter files in the repo for convenient deployment for those who already have the model installed.

No quants this time because had issues quantizing this model specifically, if you need quants you can wait until mradermacher quantizes the model.

UPD:

Nevermind, this monstrosity is seemingly cursed as mradermacher's quantization failed as well, I recommend getting the f32 gguf lora from the LoRA adapters folder in the repo and running it with the base model for convenience, you can also merge lora weights with your model if you prefer single file, atp I recommend my ministral 3 3b finetune as it's just better and didn't give me as much headache, altough this is a pretty good model if you can't go any higher than 2b, sorry if that's inconvenient to someone