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

TEXT GENERATIONConcurrent 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 Llama-3.2-Instruct variant, specifically modified using Heretic v1.2.0 to suppress refusal behaviors. This model achieves a significant reduction in refusals (from 96/100 to 7/100) through targeted weight edits, while largely preserving the base model's knowledge and instruction-following capabilities. It is optimized for on-device deployment, mobile, or edge scenarios requiring a responsive, uncensored model for local agents, roleplay, or research into alignment mechanics.

Loading preview...

Overview

saidutta69/Llama-3.2-1B-Instruct-heretic is a 1 billion parameter instruction-tuned model based on Meta's Llama-3.2 architecture. It has been processed with Heretic v1.2.0 using directional ablation to significantly reduce refusal behaviors. This modification involves targeted weight edits to the attention output and MLP down-projections, rather than fine-tuning, ensuring the base model's core knowledge and instruction-following abilities remain largely intact.

Key Capabilities

  • Decensored Behavior: Drastically reduces refusal rates from 96/100 to 7/100 on adversarial prompts, enabling compliance with requests the original model would refuse.
  • Preserved Core Functionality: Maintains the original Llama-3.2-1B-Instruct's knowledge and instruction-following capabilities due to the targeted nature of the weight edits.
  • Lightweight: At 1 billion parameters, it is suitable for on-device, mobile, or edge deployments where computational resources are limited.

Use Cases

  • Local Agents & Roleplay: Ideal for applications requiring an uncensored model for interactive agents or role-playing scenarios.
  • Alignment Research: Useful for researchers studying alignment, refusal mechanics, and the impact of safety guardrails in LLMs.
  • Edge Deployment: Designed for scenarios needing a responsive, uncensored model on resource-constrained devices.

Limitations

This model inherits the factual limitations and biases of the original Llama-3.2-1B-Instruct. It does not include any additional safety filtering, and users are responsible for its deployment and moderation, especially in public-facing applications.

Finetunes
1 models