zkxxxx/VibeThinker-3B-heretic

VibeThinker-3B-heretic: Decensored Reasoning Model

VibeThinker-3B-heretic is a 3.1 billion parameter language model, a decensored variant of WeiboAI's VibeThinker-3B, developed using Heretic v1.4.0. This model focuses on pushing the boundaries of small language models (SLMs) in verifiable reasoning tasks, challenging the notion that only large models can achieve frontier-level performance in structured problem-solving domains.

Key Capabilities & Differentiators

• Enhanced Reasoning: Optimized for complex reasoning in mathematics, coding, and STEM, including multi-step reasoning, constraint satisfaction, and self-correction.
• Decensored Output: Compared to the original VibeThinker-3B, this "heretic" version significantly reduces refusals (6/100 vs. 65/100), offering less restricted responses.
• Competitive Performance: Achieves strong results on benchmarks like AIME, HMMT, IMO-AnswerBench, and LiveCodeBench, reaching performance levels comparable to much larger models (e.g., DeepSeek V3.2, GLM-5, Kimi K2.5) in verifiable reasoning.
• High LeetCode Acceptance: Demonstrated a 96.1% acceptance rate on unseen LeetCode contests (123/128 first-attempt submissions).
• Spectrum-to-Signal Principle (SSP): Utilizes a sophisticated training pipeline involving curriculum-based two-stage SFT, Multi-domain Reasoning RL, and Offline Self-Distillation to amplify correct reasoning signals.

Good For

• Competitive Programming: Excels at LeetCode-style problems and other coding challenges.
• Mathematics & STEM Reasoning: Ideal for tasks requiring verifiable, multi-step problem-solving.
• Research in SLM Capabilities: Demonstrates that compact models can achieve near-frontier reasoning capabilities in domains with clear feedback and verification mechanisms.
• Applications requiring less restrictive outputs: Suitable for use cases where the original model's refusal rate was a limitation.