VibeThinker-3B - A 3 billion parameter intensive inference model open-sourced by Weibo.

What is VibeThinker-3B?

VibeThinker-3B is an open-source, 3 billion-parameter dense inference model developed by the Sina Weibo team. Built on Qwen2.5-Coder-3B, it achieves performance comparable to leading large models like Gemini 3 Pro and Claude Opus 4.5 on verifiable inference tasks such as mathematics and programming through an upgraded Spectrum-to-Signal post-training process. It explores the capabilities of small models and complements traditional Scaling Law.

Main functions of VibeThinker-3B

• High-difficulty mathematical reasoningIt scored 94.3 points in AIME’26, 89.3 points in HMMT’25, and 93.8 points in BruMO’25.
• Competitive programmingLiveCodeBench v6 achieved a pass rate of 80.2% with 1 pass, and LeetCode's latest weekly/bi-weekly contest pass rate is as high as 96.1%.
• STEM ReasoningIt achieves a score of 76.4 on IMO-AnswerBench (upgraded to 80.6 with +CLR), supporting the solution of complex scientific problems.
• Instructions followedIFBench scored 74.5 points, demonstrating its ability to be sensitive to formatting and constraint checks.
• Scaling during testing (CLR)Introducing a Claim-Level reliability assessment strategy can further improve the reliability of answers during the reasoning stage.

Technical Principles of VibeThinker-3B

• Base modelBuilt on Qwen2.5-Coder-3B, the 3 billion parameter-intensive architecture remains unchanged, and all performance improvements come from the extreme optimization of the post-training technology stack.
• Spectrum-to-Signal ParadigmThe core training process is divided into the "Spectrum Phase" (SFT extended capability coverage) and the "Signal Phase" (RLVR precise optimization of verifiable tasks).
• Two-stage course SFTThe first phase broadly covers mathematics, programming, STEM, and conversational skills, while the second phase focuses on challenging long reasoning samples, gradually increasing the depth of the learning process.
• Diversity exploration distillationInstead of pursuing a single optimal solution, it retains multiple valid reasoning paths, thereby enhancing the model's generalization ability on complex problems.
• MGPO Reinforcement Learning: Add extra weights to GRPO to prioritize optimizing samples that are "neither too easy nor too difficult" for the current strategy.
• Multi-domain Sequential RLReinforcement learning was performed strictly in the order of Math → Code → STEM, and experiments showed that this order yielded the best overall results.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use VibeThinker-3B

• Download ModelAccess the HuggingFace or GitHub repository to obtain the model weights.
• Local deploymentIt loads 3B parameter models based on frameworks such as transformers, and supports operation on consumer-grade hardware.
• Invoking inferenceInput a math, programming, or STEM question, and the model will output an answer with a complete chain of reasoning (CoT).
• Enable CLR (optional)For math-related tasks, you can enable Claim-Level reliability assessment to further improve accuracy.

VibeThinker-3B's core advantages

• Minimal parameters, cutting-edge performanceWith only 3B parameters, it can be benchmarked against the performance of cutting-edge models with hundreds of billions of parameters on verifiable inference tasks.
• Course-based two-stage SFTFrom broad capability coverage to challenging long reasoning samples, the depth of reasoning is gradually increased.
• Multi-domain RL enhancementReinforcement learning is performed sequentially in the order of Math → Code → STEM, preserving the complete 64K long contextual reasoning trajectory.
• Offline self-distillationHigh-quality trajectories are selected for unified distillation, prioritizing the learning of correct paths that student models have not yet mastered.

VibeThinker-3B project address

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

VibeThinker-3B Comparison with Similar Products

Application scenarios of VibeThinker-3B

• Algorithm Competition PreparationIt assists in solving math competition problems such as AIME, HMMT, and IMO, and provides multi-path reasoning references.
• Programming Interview TrainingPractice programming problems on LeetCode, LiveCodeBench, etc., and get the correct solutions with detailed explanations.
• STEM Education and TutoringIt provides a step-by-step derivation process for solving problems in science subjects such as physics, chemistry, and biology.
• Edge inference deploymentBecause the model has only 3B parameters, it is suitable for local operation in low-computing environments such as mobile phones and IoT devices.
• Research on reasoning ability: As an exploratory benchmark for the upper limit of reasoning in small models, it provides an alternative path for academic research on Scaling Law.