WeiboAI/VibeThinker-3B

VibeThinker-3B: A Small Model for Frontier Reasoning

VibeThinker-3B, developed by WeiboAI, is a 3-billion parameter model designed to excel in challenging verifiable reasoning tasks across mathematics, coding, and STEM. It builds upon the VibeThinker series, leveraging an optimized Spectrum-to-Signal Principle (SSP) post-training pipeline to achieve performance comparable to top-tier frontier reasoning models like Qwen3.6 Plus and Gemini 3 Pro, despite its compact size.

Key Capabilities & Performance

• Exceptional Reasoning: Achieves 76.4 on IMO-AnswerBench (80.6 with CLR), a highly challenging mathematical benchmark, demonstrating near-frontier performance with significantly fewer parameters than models like DeepSeek V3.2 (671B) and GLM-5 (744B).
• Strong Coding Performance: Passes 123/128 (96.1% acceptance rate) on unseen LeetCode weekly and biweekly contests (Python).
• Multi-domain Proficiency: Shows competitive results across mathematics, coding, knowledge, and instruction-following benchmarks.

Training Innovations

The model's strong performance is attributed to its advanced training pipeline, which includes:

• Curriculum-based Two-stage SFT: Progresses from broad capability coverage to harder, longer-horizon reasoning samples.
• Multi-domain Reasoning RL: Utilizes MaxEnt-Guided Policy Optimization (MGPO) applied sequentially to math, code, and STEM tasks within a 64K long-context window.
• Offline Self-Distillation: Filters and distills high-quality trajectories from RL checkpoints.

Ideal Use Cases

VibeThinker-3B is particularly recommended for:

• Competitive-style math problems.
• Coding challenges (e.g., LeetCode-style).
• STEM reasoning tasks.
• Any application where the target answer can be clearly verified.

Note: This model is not recommended for tool-calling, agent-based programming, or broad open-domain knowledge tasks, where larger general-purpose models may be more suitable.