saidutta69/VibeThinker-3B-heretic
VibeThinker-3B-heretic is a 3.1 billion parameter decensored variant of WeiboAI/VibeThinker-3B, fine-tuned from Qwen/Qwen2.5-Coder-3B, with a 32768 token context length. Developed by saidutta69, this model retains the original VibeThinker's explicit chain-of-thought reasoning and coding capabilities while having refusal guardrails removed via directional ablation. It is optimized for CPU and low-VRAM local use, agents, and on-device reasoning where uninhibited output is desired.
Loading preview...
VibeThinker-3B-heretic: Decensored Reasoning and Coding
VibeThinker-3B-heretic is a 3.1 billion parameter language model derived from WeiboAI/VibeThinker-3B, which itself is based on Qwen/Qwen2.5-Coder-3B. This variant has been processed using the Heretic v1.4.0 tool, employing "abliteration" to suppress refusal behaviors.
Key Differentiators
- Decensored Output: Refusal guardrails are removed through targeted weight edits to attention output and MLP down-projections, rather than traditional fine-tuning. This method aims to preserve the base model's original reasoning, coding, and instruction-following capabilities.
- Explicit Thinking Mode: Retains VibeThinker's unique chain-of-thought behavior, making it suitable for tasks requiring explicit reasoning steps.
- Resource-Efficient: Designed for CPU and low-VRAM environments, making it practical for local deployment, agents, and edge devices.
- Abliteration Method: Utilizes directional ablation to edit specific weights responsible for refusal, ensuring the core network capabilities remain largely intact. More details on this technique can be found in the Heretic repo and the original writeup.
Ideal Use Cases
- Users requiring a ~3B parameter model for reasoning and coding tasks without any refusal guardrails.
- Applications on CPU-only systems or devices with limited VRAM.
- Development of AI agents where unconstrained output is a requirement.
- On-device reasoning applications where a compact, capable model is needed.