WeiboAI, the research lab inside Chinese social platform Weibo, released VibeThinker-3B, a 3-billion-parameter model built on Qwen2.5-Coder-3B and tuned for verifiable reasoning, meaning problems with a checkable answer like math and code. It scores 94.3 on AIME26, a hard math competition, 80.2 on LiveCodeBench v6 (solved on the first try), and a 96.1% acceptance rate on LeetCode contests it had never seen. The technical report shows it matches or beats DeepSeek V3.2, GLM-5, and Gemini 3 Pro, models hundreds of times larger. A test-time method that spends extra compute checking each step lifts AIME26 to 97.1. A 93.4 on IFEval confirms the reasoning push did not break instruction following.
Frontier-level math and coding now fit on a laptop instead of a server rack. The earlier 1.5B version cost just $7,800 to post-train, against $294K for DeepSeek R1 and $535K for MiniMax-M1, a 30x to 60x drop. A student, a small startup, or a hospital can run a top reasoning model locally, cheaply, and privately, with weights released under the permissive MIT license.
The team argues reasoning packs into a small core, while broad world knowledge still needs scale. That points to two model sizes for two jobs, not one giant model for everything. It echoes NVIDIA’s Nemotron-Cascade 2, which hit math-olympiad gold with 3B active parameters.
Sources:
- VibeThinker-3B Technical Report (arXiv)
- VibeThinker-3B on Hugging Face
- VibeThinker GitHub Repository
- Technical Report on Hugging Face
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {VibeThinker-3B: {Frontier} {Math} and {Coding} {Reasoning} in
a {3B} {Model}},
date = {2026-06-29},
url = {https://toknow.ai/posts/vibethinker-3b-frontier-verifiable-reasoning-small-model/},
langid = {en-GB}
}