VibeThinker-3B is a three-billion-parameter open model from Weibo parent company Sina that matches far larger systems on math and coding benchmarks, sometimes rivaling models hundreds of times its size. The gains come not from scale but from a carefully staged post-training pipeline applied to an existing base model, challenging the assumption that bigger is always better.
Where the small model shines, and where it doesn't
On structured tasks with verifiable solutions, such as math olympiads and programming challenges, VibeThinker-3B performs on par with vastly larger systems. On the AIME'26 benchmark it matches DeepSeek V3.2 (around 671B parameters) and Kimi K2.5 (about 1T parameters), and on IMO-AnswerBench it scores 76.4, rising to 80.6 with an additional consistency step, comparable to those giants. On LiveCodeBench v6 it beats every open-source model under 20 billion parameters.
To rule out data contamination, the team entered VibeThinker-3B in live LeetCode contests held after training finished, between April and May 2026. There it solved 123 of 128 problems on the first attempt and outscored much larger systems including GPT-5.2, Qwen3-Max, and Kimi K2.5. That fresh, unseen test is strong evidence the results are not the product of memorized answers.
Factual knowledge is the clear weak spot. On the knowledge-heavy GPQA-Diamond benchmark, the model falls well behind its much larger rivals, which can store far more facts in their parameters. Follow more model releases in our AI section.
- Parameters: 3 billion (built on Alibaba's Qwen2.5-Coder-3B)
- Strength: math, coding, and structured reasoning
- Weakness: broad factual recall (GPQA-Diamond)
- Availability: open on Hugging Face and GitHub
How the training pipeline works
Sina's real contribution is the post-training recipe rather than the architecture. The model first learns broadly through two stages of supervised fine-tuning across math, coding, and general dialogue. It is then tailored for hard, multi-step reasoning through reinforcement learning applied in sequence, first for math, then programming, then STEM.
A self-distillation step consolidates those skills back into a single coherent model, and a final reinforcement-learning pass sharpens instruction-following. Stacking specialized training phases this way lets a tiny model punch far above its parameter count on the narrow, verifiable tasks the pipeline targets.
What it means for AI capabilities
The authors frame their results around a "Parametric Compression-Coverage Hypothesis." The idea is that logical reasoning relies on a relatively small set of recurring patterns and can be packed into a compact core, while broad world knowledge needs many parameters simply to store many facts. In that view, reasoning compresses well and knowledge does not.
That reframes small models as an independent research direction rather than budget versions of large ones. Where tasks are clearly structured and answers can be verified, a compact, cheap-to-run model may be all that is needed, while encyclopedic recall still favors scale. Small models catching up on narrow tasks is becoming a recurring pattern across the field.
The practical upside is access. An open three-billion-parameter model that rivals trillion-parameter systems on math and code can run on far cheaper hardware, lowering the barrier for researchers, startups, and hobbyists. Read more analysis in our science coverage.
VibeThinker-3B: quick FAQ
How big is VibeThinker-3B compared to the models it matches?
At three billion parameters, it rivals systems such as DeepSeek V3.2 and Kimi K2.5 that are hundreds of times larger on math and coding benchmarks, though only on those structured tasks.
Where can I download VibeThinker-3B?
It is released openly on Hugging Face and GitHub, and because of its small size it can run on far less hardware than trillion-parameter rivals.
What is VibeThinker-3B not good at?
Broad factual recall. On knowledge-heavy tests like GPQA-Diamond it trails much larger models, reflecting the trade-off described in its Parametric Compression-Coverage Hypothesis.
