What is VibeThinker-3B?
VibeThinker-3B is a small reasoning model from Weibo's AI team, released in June 2026 as the successor to VibeThinker-1.5B. It is a 3-billion-parameter dense model fine-tuned from Alibaba's Qwen2.5-Coder-3B, and it reads and writes text. Weibo built it for problems with checkable answers: competition maths, competitive programming and science questions.
Training ran in four stages: two rounds of supervised fine-tuning moving from broad to harder, longer reasoning; reinforcement learning on maths, code and science in turn, with a 64K-token window; distilling the best traces back into one model; and a final round to improve instruction following. Weibo frames it as a test of a hypothesis that verifiable reasoning needs far fewer parameters than broad knowledge does.
The weights are under the MIT licence. In the last 30 days they drew about 2,800 Hugging Face downloads.
Is VibeThinker-3B worth it?
VibeThinker-3B's maths and contest-coding scores are startling for 3 billion parameters. Weibo's technical report gives 94.3 on AIME 2026, or 97.1 with its claim-level test-time checking, 80.2 on LiveCodeBench v6 and 93.4 on IFEval. On IMO-AnswerBench's 400 olympiad problems it scores 76.4, rising to 80.6 with checking, against 78.3 for the 671B DeepSeek V3.2 and 82.5 for the 744B GLM-5.
The best evidence that this is not just benchmark tuning is the LeetCode test. On weekly and biweekly contests from 25 April to 31 May 2026, which Weibo says the model had not seen, it passed 123 of 128 problems first time, a 96.1% acceptance rate. That is still Weibo's own measurement.
The catch is everything outside that lane. Weibo's card says it was not trained on tool calling or agent data and advises against using it for function calling, API orchestration or coding agents, and says larger models remain better for broad knowledge. Hard problems need very long outputs: Weibo suggests 60,000 to 100,000 tokens for its hardest maths set, and its sample code allows 102,400 new tokens, which is slow on the laptops a 3B model invites.
It is best treated as a specialist solver. For a general on-device assistant, Nanbeige4.2-3B or Spark-X2.6-4B are built for tool use and agents; for a self-hosted maths or contest-programming checker, its scores are hard to match at 3B, and the 1B Limite Violetto is the rival to compare it with, since it beats VibeThinker on several of the same tests.
Where it earns its place
A maths teacher who coaches a school olympiad team
Generate worked solutions to past AIME and olympiad problems for practice sets
Run it with vLLM on a single consumer GPU at temperature 1.0 and allow long outputs. Its 94.3 on AIME 2026 suits this; check each solution, since the score is self-reported.
An engineer at an online coding-practice platform
Draft reference solutions for new algorithm problems before human review
Batch new problems through SGLang and keep solutions that pass your test cases. Weibo reports 123 of 128 recent LeetCode contest problems passed first time; the MIT licence allows commercial use.
Before you commit
- Setup
- Install transformers 4.54.0 or later and load in bfloat16, or serve with vLLM 0.10.1 or SGLang 0.4.9.post6. Weibo's evaluation settings are temperature 1.0 and top_p 0.95.
- Integrations
- transformers, vLLM and SGLang; ModelScope mirror. No tool-calling template is provided.
- Data and privacy
- Self-hosted, so problems and code stay on your hardware. MIT licence permits commercial use and modification.
- Cost at scale
- Weights are free and a 3B model fits a single consumer GPU, but long reasoning traces of tens of thousands of tokens per hard problem dominate compute time.
- Limits
- Not suited to function calling, agents or open-domain knowledge; very long outputs on hard problems; benchmarks are Weibo's own.
What’s good
- 94.3 on AIME 2026 and 80.2 on LiveCodeBench v6 (self-reported)
- 96.1% first-attempt pass rate on unseen LeetCode contests
- MIT licence, standard Qwen2 architecture
- Runs on vLLM and SGLang with no custom code
What’s not
- Not trained for tool calling or agents
- Weak on broad knowledge by Weibo's own account
- Hard problems need 60,000 to 100,000 output tokens
- Text only, English-focused
Best for
Teams that need a cheap, self-hosted solver or checker for competition maths, science questions or algorithm problems.
VibeThinker-3B alternatives
VibeThinker-3B: common questions
- Is VibeThinker-3B worth it?
- VibeThinker-3B solves olympiad maths and contest code like a model hundreds of times larger, and should be kept well away from anything else. We rate it 3.7 out of 5.
- How much does VibeThinker-3B cost?
- Free MIT-licensed weights. No first-party hosted API published.
- What are the downsides of VibeThinker-3B?
- Not trained for tool calling or agents; weak on broad knowledge by Weibo's own account; hard problems need 60,000 to 100,000 output tokens; text only, English-focused.
- Who is VibeThinker-3B best for?
- Teams that need a cheap, self-hosted solver or checker for competition maths, science questions or algorithm problems.
- How much does VibeThinker-3B cost for a team or in production?
- Weights are free and a 3B model fits a single consumer GPU, but long reasoning traces of tens of thousands of tokens per hard problem dominate compute time.
- Is VibeThinker-3B safe for company data?
- Self-hosted, so problems and code stay on your hardware. MIT licence permits commercial use and modification.
- What are the best VibeThinker-3B alternatives?
- The closest alternatives we have reviewed are Nanbeige4.2-3B, Spark-X2.6-4B, Limite 1B - Violetto and gpt-oss-20b.
Answers checked Nov 7, 2026.