VibeThinker-3B - Weibo's Open-Source 3-Billion-Parameter Dense Reasoning Model

What is VibeThinker-3B

VibeThinker-3B is a 3-billion-parameter dense reasoning model open-sourced by the Sina Weibo team, built on Qwen2.5-Coder-3B and, through an upgraded Spectrum-to-Signal post-training pipeline, achieves Gemini 3 Pro、Claude Opus 4.5 and other frontier large models' performance on verifiable reasoning tasks such as mathematics and programming, exploring the capability boundaries of small models and complementing the traditional Scaling Law.

VibeThinker-3B's Main Features

• High-Difficulty Mathematical Reasoning: Achieves 94.3 points on AIME'26, 89.3 points on HMMT'25, and 93.8 points on BruMO'25.
• Competitive Programming: Achieves 80.2 Pass@1 on LiveCodeBench v6, with a pass rate of up to 96.1% on the latest LeetCode weekly/biweekly contests.
• STEM Reasoning: Achieves 76.4 points on IMO-AnswerBench (+CLR improves it to 80.6), supporting complex scientific problem solving.
• Instruction Following: Achieves 74.5 points on IFBench, with format-sensitive and constraint-checking capabilities.
• Test-Time Scaling (CLR): Introduces a Claim-Level reliability evaluation strategy that can further improve answer reliability during the inference stage.

VibeThinker-3B's Technical Principles

• Base Model: Built on Qwen2.5-Coder-3B, with the 3-billion-parameter dense architecture unchanged; all performance improvements come from extreme optimization of the post-training technology stack.
• Spectrum-to-Signal Paradigm: The core training pipeline is divided into a "spectrum stage" (SFT expanding capability coverage) and a "signal stage" (RLVR precisely optimizing verifiable tasks).
• Two-Stage Curriculum SFT: The first stage broadly covers mathematics, programming, STEM, and dialogue capabilities, while the second stage focuses on high-difficulty long-reasoning samples, progressively increasing depth.
• Diversity Exploration Distillation: Rather than pursuing a single optimal solution, it preserves multiple effective reasoning paths, enhancing the model's generalization ability on complex problems.
• MGPO Reinforcement Learning: Adds extra weight on top of GRPO, prioritizing samples that are "neither too easy nor too hard" for the current policy during optimization.
• Multi-Domain Sequential RL: Reinforcement learning is conducted strictly in the order Math → Code → STEM; experiments found this order to be optimal for overall performance.

Follow WeChat and reply "Open Source" to join theAI Open-Source Project Exchange Group

How to Use VibeThinker-3B

• Download the Model: Visit the HuggingFace or GitHub repository to obtain the model weights.
• Local Deployment: Load the 3B-parameter model based on frameworks such as transformers, supporting operation on consumer-grade hardware.
• Invoke Inference: Input mathematics, programming, or STEM questions, and the model will output answers with a complete chain of thought (CoT).
• Enable CLR (Optional): For mathematical tasks, Claim-Level reliability evaluation can be enabled to further improve accuracy.

VibeThinker-3B's Core Advantages

• Extremely Small Parameters, Frontier Performance: With only 3B parameters, it can match the performance of hundred-billion-scale frontier models on verifiable reasoning tasks.
• Curriculum-Style Two-Stage SFT: From broad capability coverage to high-difficulty long reasoning samples, gradually increasing reasoning depth.
• Multi-domain RL reinforcement: Perform reinforcement learning sequentially in the order Math → Code → STEM, retaining complete 64K long-context reasoning trajectories.
• Offline self-distillation: Filter high-quality trajectories for unified distillation, prioritizing learning the correct paths that the student model has not yet mastered.

VibeThinker-3B project address

• GitHub repository：https://github.com/WeiboAI/VibeThinker
• HuggingFace model library：https://huggingface.co/WeiboAI/VibeThinker-3B
• arXiv technical paper：https://arxiv.org/pdf/2606.16140

Comparison of VibeThinker-3B with similar competing products

Application scenarios of VibeThinker-3B

• Algorithm competition preparation: Assist in solving math competition problems such as AIME, HMMT, IMO, etc., providing multi-path reasoning references.
• Programming interview training: Practice programming problems such as LeetCode, LiveCodeBench, etc., and obtain correct solutions with detailed explanations.
• STEM education tutoring: Provide step-by-step derivation of solution processes for science problems such as physics, chemistry, and biology.
• Edge inference deployment: Because the model has only 3B parameters, it is suitable for local operation in low-compute environments such as mobile phones and IoT devices.
• Reasoning capability research: As an exploratory benchmark for the upper limit of small-model reasoning, it provides academia with an alternative path for studying Scaling Law.