WeiboAI

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 Qwen2.5-Coder-3B, whose own Qwen Research License is not OSI-approved, released without the training data or pipeline, so open weights rather than open source.

• https://huggingface.co/WeiboAI/VibeThinker-3B recorded 2026-08-13

  `license:mit` in the card data and `"gated":false`; `base_model - Qwen/Qwen2.5-Coder-3B`; no RL/SFT mixture and no training pipeline published

Adoption is measured as Hugging Face downloads on VibeThinker-3B, the product's one repository.

• https://huggingface.co/api/models/WeiboAI/VibeThinker-3B recorded 2026-08-14

  149,008 downloads in the trailing 30 days for WeiboAI/VibeThinker-3B, ungated

• https://huggingface.co/WeiboAI/VibeThinker-3B recorded 2026-08-13

  active HF repo; 822 likes; early traction as a small reasoning specialist

VibeThinker-3B reaches frontier-range results on narrow, verifiable math and code reasoning despite having only 3B parameters, according to its model card. It is explicitly not built for general or agentic use, so it sits below general-purpose frontier models.

• https://huggingface.co/WeiboAI/VibeThinker-3B recorded 2026-09-29

  Model page embeds eval results for MathArena/aime_2026 and Idavidrein/gpqa (diamond) with value 70.2; the page states "VibeThinker-3B reaches 76.4 on IMO-AnswerBench".

• https://huggingface.co/WeiboAI/VibeThinker-3B/raw/main/README.md recorded 2026-09-29

  Model card: "VibeThinker-3B reaches 76.4 on IMO-AnswerBench, a highly challenging benchmark with 400 IMO-level problems, with only 3B parameters"; recent LeetCode contests "corresponding to a" 96.1% acceptance rate; warns it "was not trained on tool-calling or agent-based programming data".