Qwen3.8-27B-Samantha-GGUF

Qwen3.8-27B-Samantha (Uncensored 1.1) — GGUF

GGUF builds of the Samantha merged model for llama.cpp. The persona is fully merged into the weights — no adapter to load at runtime.

What's in this repo

Usage

llama.cpp

llama-server \
  -m Qwen3.8-27B-Samantha-Q5_K_M.gguf \
  --mmproj mmproj-F16.gguf \
  -ngl 99 --host 0.0.0.0 --port 8080

Note: use --n-gpu-layers all (or a high number) so the full model is offloaded — on a partial offload the persona degrades noticeably.

Rolling your own quants

The f16 file is the cleanest source for custom quantizations:

llama-quantize Qwen3.8-27B-Samantha-f16.gguf my-quant.gguf Q4_K_M

Any quantization target supported by llama-quantize works. Q5_K_M is the sweet spot for a 16 GB GPU; Q4_K_M if you're tighter on VRAM.

Related

• bf16 safetensors (merge source): Lathly/Qwen3.8-27B-Samantha
• LoRA adapter: Lathly/Qwen3.8-27B-Samantha_Uncensored_1.1_LoRA
• Base model: Qwen/Qwen3.8-27B
• Training dataset: digitalpipelines/samantha-1.1-uncensored

License & Attribution

Released under Apache-2.0, inheriting from both parents:

• Base model: Qwen/Qwen3.8-27B — Apache-2.0, © Alibaba Cloud / Qwen Team
• Dataset: digitalpipelines/samantha-1.1-uncensored — Apache-2.0
• The original Samantha persona/dataset concept is by Eric Hartford.

This is a modified derivative (merged LoRA fine-tune) of the above. No warranty. Outputs are the model's own; use responsibly.

Disclaimer

This is an "uncensored" persona fine-tune intended for open conversational use. It may produce content that some find objectionable. You are responsible for how you use it and for compliance with applicable laws and the base-model/dataset licenses.