saidutta69/Qwen2.5-14B-Instruct-heretic

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 16, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

saidutta69/Qwen2.5-14B-Instruct-heretic is a 14.8 billion parameter instruction-tuned causal language model derived from Qwen/Qwen2.5-14B-Instruct. Developed by RACER IS OP, this variant has its refusal behaviors suppressed using the Heretic v1.4.0 abliteration method, which involves targeted weight edits rather than fine-tuning. This model maintains the base model's strong reasoning and instruction-following capabilities while providing direct answers, making it suitable for developers requiring uncensored responses.

Loading preview...

Model Overview

saidutta69/Qwen2.5-14B-Instruct-heretic is a 14.8 billion parameter model based on Qwen/Qwen2.5-14B-Instruct. Its primary distinction is the removal of refusal behaviors through a technique called abliteration, using Heretic v1.4.0. This method directly edits specific weights responsible for refusal in the attention output and MLP down-projections, preserving the base model's core knowledge and instruction-following abilities.

Key Differentiators

  • Decensored Responses: Significantly reduces refusal rates from 98/100 to 14/100 on adversarial prompts, ensuring direct answers.
  • Capability Preservation: Unlike fine-tuning, abliteration maintains the original model's coherence and capabilities by making narrow, targeted edits, resulting in a low KL divergence of 0.0600 from the base model.
  • Hardware Accessibility: Provided with a full GGUF quantization ladder (Q4_K_M, Q5_K_M, Q6_K, Q8_0), enabling efficient deployment on consumer GPUs ranging from 6GB to 24GB VRAM, as well as CPU-only setups.

Use Cases

This model is designed for developers who require a powerful instruction-following LLM that provides direct, uncensored responses without the refusal guardrails present in the base Qwen2.5-14B-Instruct. It is ideal for applications where unfiltered output is necessary, provided the user manages responsible deployment. It is not a capability upgrade but a behavioral modification of the base model.