saidutta69/Qwen3-1.7B-heretic

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Qwen3-1.7B-heretic is a 1.7 billion parameter, 32k context length language model developed by saidutta69, based on the Qwen3 architecture. This model is a decensored variant of Qwen/Qwen3-1.7B, produced using the Heretic v1.4.0 method which suppresses refusal behavior via targeted weight edits rather than fine-tuning. It is primarily designed for developers seeking an uncensored Qwen3 base model for experimentation, fine-tuning, and studying refusal mechanics, while retaining the original model's knowledge and hybrid thinking capabilities.

Loading preview...

Overview

Qwen3-1.7B-heretic is a 1.7 billion parameter language model derived from the Qwen3 architecture, developed by saidutta69. This model is a decensored variant of the original Qwen/Qwen3-1.7B, created using the Heretic v1.4.0 tool. Unlike traditional fine-tuning, Heretic employs directional ablation (abliteration) with per-layer refusal directions, specifically targeting attention output and MLP down-projections to suppress refusal behavior. This method ensures that the base model's core knowledge and "hybrid thinking capability" are largely preserved.

Key Capabilities

  • Decensored Output: Significantly reduced refusal behavior compared to the base model (4/100 refusals vs. 92/100 for the original Qwen3-1.7B).
  • Preserved Knowledge: Retains the original Qwen3-1.7B's knowledge and reasoning abilities due to the targeted weight editing approach.
  • Accessibility: Small enough (1.7B parameters) to run efficiently on consumer hardware, including CPU-only machines, via GGUF quantizations.
  • Experimentation Platform: Ideal for developers interested in uncensored model experimentation, serving as a base for further fine-tuning, or studying the mechanics of refusal in LLMs.

Good for

  • Uncensored Applications: Developing applications where strict refusal guardrails are undesirable or need to be customized.
  • Research & Development: Investigating model behavior, safety mechanisms, and the impact of refusal suppression techniques.
  • Resource-Constrained Environments: Deploying a capable language model on gaming PCs or other machines with limited GPU memory, thanks to comprehensive GGUF support.
  • Fine-tuning Base: Providing a clean, uncensored foundation for further domain-specific fine-tuning without inherent refusal biases.

Finetunes
1 models