saidutta69/VibeThinker-3B-heretic
VibeThinker-3B-heretic by saidutta69 is a 3.1 billion parameter decensored variant of WeiboAI/VibeThinker-3B, fine-tuned from Qwen/Qwen2.5-Coder-3B. This model utilizes 'abliteration' via targeted weight edits to suppress refusal behavior, preserving the base model's reasoning, coding, and instruction-following capabilities. It is optimized for CPU and low-VRAM local use, agents, and edge/on-device reasoning, offering an explicit thinking mode without refusal guardrails.
Loading preview...
Overview
VibeThinker-3B-heretic is a 3.1 billion parameter language model developed by saidutta69, derived from WeiboAI/VibeThinker-3B, which itself is fine-tuned from Qwen/Qwen2.5-Coder-3B. This model's key differentiator is its 'decensored' nature, achieved through a technique called abliteration (directional ablation) using Heretic v1.4.0. Unlike traditional fine-tuning, abliteration directly edits specific weights responsible for refusal behaviors, leaving the model's core reasoning, coding, and instruction-following capabilities largely intact.
Key Capabilities
- Decensored Output: Refusal guardrails are suppressed, allowing for broader content generation.
- Reasoning and Coding: Retains the strong reasoning and coding abilities of its base model, VibeThinker-3B.
- Explicit Thinking Mode: Features VibeThinker's chain-of-thought behavior for enhanced reasoning.
- Instruction Following: Maintains robust instruction-following capabilities.
Good For
- Local Deployment: Suitable for CPU and low-VRAM environments.
- AI Agents: Ideal for integration into autonomous agents requiring unconstrained responses.
- Edge/On-Device Reasoning: Designed for efficient performance in resource-limited settings.
- Users requiring models without refusal guardrails: For applications where the base model's refusal behavior is undesirable.