VibeThinker-3B
WeiboAIWeiboAI's VibeThinker-3B, an MIT-licensed 3B reasoning model fine-tuned from Qwen2.5-3B / Qwen2.5-Coder-3B and specialized for verifiable math, coding and STEM. Claims frontier-range scores on narrow reasoning benchmarks despite its size; not intended for tool-calling or agentic use.
MIT-licensed weights, but a fine-tune of the open-weight Qwen2.5 line without a fully open data/training pipeline -> open_weights, not open_source. Narrow specialist.
Openness
3 medium confidence- weights
- open(MIT)
- base
- open_weights(Qwen2.5)
- data
- partial(RL/SFT data not fully released)
- license
- MIT
MIT-licensed weights, but a fine-tune of the open-weight Qwen2.5 line without a fully open/reproducible data+training pipeline -> open_weights tier, not open_source.
- https://huggingface.co/WeiboAI/VibeThinker-3B recorded 2026-06-30
VibeThinker-3B; MIT license; fine-tuned from Qwen2.5-3B / Qwen2.5-Coder-3B
Adoption
2 low confidenceRecently released specialist model with early attention for its size; directional estimate (not a measured usage figure).
- https://huggingface.co/WeiboAI/VibeThinker-3B recorded 2026-06-30
active HF repo; early traction as a small reasoning specialist
Capability
3 medium confidenceFrontier-range on narrow verifiable math/code reasoning despite 3B, but explicitly not for general/agentic use -> mid-tier on a general-capability scale.
- https://huggingface.co/WeiboAI/VibeThinker-3B recorded 2026-06-30
reports AIME 2026 94.3, GPQA Diamond 70.2, IMO-AnswerBench 76.4; not for tool/agent use
Unchanged since 2026-06-30 (last edited, not re-checked)