s3nh/Bonsai-27B-unpacked-abliterated-uncensored

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

s3nh/Bonsai-27B-unpacked-abliterated-uncensored is a 27 billion parameter, 32768-token context length language model, derived from prism-ml/Bonsai-27B-unpacked. This version has been decensored using Heretic v1.4.0, significantly reducing refusal rates compared to the original model. It is provided in FP16 Safetensors format for compatibility with standard HuggingFace tooling, though it does not retain the memory and speed advantages of the native 1-bit Bonsai models.

Loading preview...

Model Overview

s3nh/Bonsai-27B-unpacked-abliterated-uncensored is a 27 billion parameter language model with a 32768-token context length. It is a decensored variant of the original prism-ml/Bonsai-27B-unpacked model, processed using Heretic v1.4.0. This "abliterated" version aims to reduce content refusals.

Key Characteristics

  • Decensored Variant: Significantly reduces content refusals, showing 6 refusals per 100 prompts compared to 81/100 in the original model.
  • FP16 Safetensors: Provided in full-size FP16 format for compatibility with standard HuggingFace frameworks and tooling.
  • Unpacked Version: This model is an unpacked version of the 1-bit Bonsai 27B, meaning it does not offer the memory reduction or speed benefits of the native 1-bit models. It occupies approximately 54 GB.

Usage Considerations

This model is intended for users who require a decensored version of Bonsai-27B and need compatibility with stock HuggingFace tooling. For optimal performance, memory efficiency, and speed, it is strongly recommended to use the native 1-bit Bonsai models, such as Bonsai-27B-mlx-1bit for Apple Silicon or 1-bit GGUF for llama.cpp, which offer substantial memory savings and faster inference.

Finetunes
1 models