3B Open Model That Crushes Competitors in Math

This 3 billion parameter model performs similarly to 671 billion and 1 trillion parameter models on difficult math problems. VibeThinker-3B, developed by Sina Weibo and licensed under the permissive MIT License, achieved a score of 94.3 on the AIME26 math test, surpassing DeepSeek V3.2 (671 billion parameters) with a score of 94.2 and Kimi K2.5 (1 trillion) with a score of 93.3. It solved 123 out of 128 questions in the first try on the test set of unsolved LeetCode competitions.

It’s small, but it’s mighty. VibeThinker-3B is a post-trained model derived from Qwen2.5-Coder-3B, with BF16 parameters taking up about 6GB of space. That means it can be deployed to a consumer-level GPU. This tiny model is going head to head with much larger models on a particular task.

1How a small model punches up

How? The magic lies in their training methodology, dubbed Spectrum-to-Signal. To put it simply, first train your model to understand a variety of approaches to solving the problem (“spectrum”), and then refine the best answers using reinforcement learning to focus on problems near the limits of the model’s abilities. At runtime, you can optionally generate multiple solutions, verify the correctness of its reasoning, and take a vote to arrive at the most trustworthy response, achieving a peak score of 97.1 on AIME26 without any additional parameters.

Let’s not get carried away. While VibeThinker-3B holds its own against giant models on math and coding competitions, it falls short on general knowledge tasks like GPQA-Diamond (70.2 vs DeepSeek 82.4, Kimi 87.6). Factual knowledge does indeed follow a law of diminishing returns, and there are only so many things a 3B parameter model can remember. This is a scalpel, not a Swiss army knife.

2Why it matters for you

So what does this mean? You should be able to deploy a 3B parameter model locally and privately for the cost of electricity rather than paying a premium API provider per token. If you’re building an AI-powered product, this means the difference between a free feature and a costly disaster.

It also pokes holes in the idea that more is always better. Rather than thinking of 2026 as a competition between trillion parameter labs, think of it as a world where specialized small models are good enough that you have a choice. VibeThinker-3B is MIT licensed, meaning you can freely download, inspect, finetune, and deploy it without asking for permission, worrying about usage limits, or fearing a shutdown button.

What I’m trying to say is, don’t default to the largest model. Often times, a well-tuned small model is faster, cheaper, and completely under your control, which is priceless when dealing with APIs you don’t own and who may raise prices or yank your model.