VibeThinker 3B
WeiboAI/VibeThinker-3B
Open Source · chat · open-weights · seen 7m ago
Context
64K
Max output
—
Input $/1M
—
Output $/1M
—
Modalities
text
Released
12 Jun 2026
License: mit · WeiboAI/VibeThinker-3B
AI summary
● machine-written
VibeThinker-3B: 3B reasoning model with 64K context, MIT license
VibeThinker-3B is a 3-billion-parameter reasoning model built on Qwen2.5-Coder-3B using a Spectrum-to-Signal post-training pipeline. Trained with supervised fine-tuning, reinforcement learning, and self-distillation, it achieves benchmark scores comparable to models hundreds of times larger on verifiable math, code, and STEM tasks. The model is released under MIT license with open weights and runs on standard GPU stacks.
What's new
- Scores 94.3 on AIME26, 80.2 Pass@1 on LiveCodeBench v6, 70.2 on GPQA-Diamond
- Achieves 96.1% acceptance rate on unseen LeetCode problems (Apr–May 2026)
- Uses 64K context window throughout training; drops progressive context expansion
- CLR test-time scaling adds no parameters, improves AIME26 to 97.1
Best for
Mathematical reasoning and verificationCode generation and competitive programmingSTEM problem-solving with verifiable answersReasoning tasks requiring long context up to 64K tokens