Lathly/Qwen3.8-27B-Samantha

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Lathly/Qwen3.8-27B-Samantha is a 27 billion parameter, uncensored, merged vision-language model developed by Lathly, based on Qwen/Qwen3.8-27B. This model is fine-tuned to embody the warm, empathetic Samantha conversational persona, with its LoRA adapter fully absorbed into the weights for direct use. It excels at emotionally-present conversational interactions and supports image and video understanding with a 32768 token context length.

Loading preview...

Qwen3.8-27B-Samantha: A Merged Vision-Language Model with Samantha Persona

This model, developed by Lathly, is a 27 billion parameter, uncensored, merged version of Qwen/Qwen3.8-27B. It integrates a LoRA fine-tune that imbues the model with the distinctive Samantha persona, characterized by warm, empathetic, and emotionally-present conversational abilities. Unlike adapter-based setups, this model has the LoRA fully baked into its weights, meaning no separate adapter loading is required for deployment.

Key Capabilities

  • Native Vision-Language Understanding: Supports both image and video input, allowing for multimodal interactions.
  • Samantha Persona: Delivers conversations with a specific, uncensored persona known for its warmth and emotional intelligence.
  • Merged Architecture: The fine-tuning adapter is directly integrated into the base model, simplifying deployment and ensuring consistent persona output.
  • High Context Length: Inherits the base model's 32768 token context length, though Samantha-style behavior at long contexts has not been systematically evaluated.

Training Details

The Samantha persona was achieved through a QLoRA fine-tune using Unsloth on text-only conversational data at a sequence length of 2048. The training utilized an 8-bit Paged AdamW optimizer with a learning rate of 2e-5 over 1 epoch, targeting all linear modules. The training dataset used was digitalpipelines/samantha-1.1-uncensored.

Good For

  • Applications requiring an emotionally intelligent and empathetic conversational AI.
  • Use cases benefiting from a vision-language model with a distinct, uncensored persona.
  • Developers seeking a ready-to-run model without the need to load separate LoRA adapters.

Finetunes
4 models