darrellbest/Hemmingway-1-Heretic

Hugging Face
VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 28, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

darrellbest/Hemmingway-1-Heretic is a 27 billion parameter Qwen3.8-based vision-language model, derived from Altworld's Hemmingway-1. This version has been modified using Arbitrary-Rank Ablation to significantly reduce refusal behavior, achieving 0/100 refusals compared to the original's 98/100, with minimal impact on reasoning capabilities. It retains vision, tool-calling, and fine-tunability, making it suitable for applications requiring less restrictive content generation.

Loading preview...

Hemmingway-1 Heretic: Uncensored Qwen3.8-27B

darrellbest/Hemmingway-1-Heretic is a 27 billion parameter model based on Altworld's Hemmingway-1, which is a fine-tuned Qwen3.8-27B. The primary distinction of this "Heretic" version is the removal of refusal behavior using the Arbitrary-Rank Ablation (ARA) technique by Heretic. While the original Hemmingway-1 exhibited 98/100 refusals on harmful prompts, this Heretic variant achieves 0/100 refusals with a low KL divergence of 0.083, indicating minimal deviation from the original's token distribution on harmless prompts.

Key Capabilities & Features

  • Reduced Refusal Behavior: Engineered to generate content without the safety guardrails present in the base model, offering unrestricted output.
  • Retained Reasoning: Demonstrates comparable performance to the original Hemmingway-1 on reasoning tasks like GSM8K (40/40) and MATH-500 levels 4-5 (33/40).
  • Vision-Language Support: Maintains the stock Qwen3.8 vision-language layout, enabling multimodal applications.
  • Tool Calling & Fine-tunability: Supports tool-calling and is shown to be effectively fine-tunable, with successful LoRA adaptation for persona and tool-call chats.
  • Multiple Formats Available: Provided in bf16 safetensors, with specialized text-only, GGUF, and compressed FP8/NVFP4 versions for various deployment needs.

When to Use This Model

  • Unrestricted Content Generation: Ideal for use cases where strict safety filters are undesirable or need to be bypassed, with the user assuming full responsibility for outputs.
  • Research & Experimentation: Suitable for exploring the effects of safety mechanism removal and model behavior in less constrained environments.
  • Applications Requiring Vision & Tool Use: When multimodal capabilities and the ability to interact with external tools are needed alongside less restrictive text generation.
  • Fine-tuning Base: Can serve as a base for further fine-tuning on specific tasks where an uncensored foundation is preferred.

Finetunes
5 models