saidutta69/Llama-3.2-1B-Instruct-heretic

TEXT GENERATIONPricing:Input $0.108 / Output $0.804Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 2, 2026License:llama3.2Architecture:Transformer Featherless Exclusive Cold

Llama-3.2-1B-Instruct-heretic by saidutta69 is a 1 billion parameter instruction-tuned causal language model, a decensored variant of Meta's Llama-3.2-1B-Instruct. It suppresses refusal behavior through targeted weight edits using Heretic v1.2.0, maintaining the base model's knowledge and instruction-following. This model is optimized for on-device, mobile, or edge deployments requiring a responsive uncensored model, ideal for local agents, roleplay, or research into alignment mechanics.

Loading preview...

Overview

Llama-3.2-1B-Instruct-heretic is a 1 billion parameter instruction-tuned model derived from meta-llama/Llama-3.2-1B-Instruct. Developed by RACER IS OP using the Heretic v1.2.0 tool, this variant specifically targets and suppresses refusal behaviors through directional ablation (weight edits to attention output and MLP down-projections) rather than fine-tuning. This method ensures that the base model's core knowledge and instruction-following capabilities remain largely intact.

Key Capabilities & Differentiators

  • Decensored Behavior: Significantly reduces refusal rates (from 96/100 to 7/100 adversarial prompts) compared to the original Llama-3.2-1B-Instruct, while maintaining a low KL divergence of 0.1713, indicating minimal deviation from the base model's original capabilities.
  • Efficient Deployment: At 1 billion parameters, it is well-suited for on-device, mobile, or edge computing scenarios, offering responsiveness for local applications.
  • Hardware Compatibility: Provides a full suite of GGUF quantizations, enabling efficient execution on various consumer GPUs (e.g., RTX 3060/4070 with Q8_0, RTX 4060/3070 with Q6_K) and even CPU-only or Apple Silicon setups.

Ideal Use Cases

This model is particularly useful for developers and researchers who require:

  • Uncensored Agents: Building local agents or applications where the base model's refusal guardrails are undesirable.
  • Roleplay & Creative Applications: Scenarios demanding less restrictive content generation.
  • Alignment Research: Investigating refusal mechanics and alignment strategies in LLMs.
  • Edge & Mobile Deployment: Deploying a capable, uncensored model on resource-constrained devices.

Finetunes
1 models