A 3B Open Model That Rivals Giants at Math

A new open-source model with just 3 billion parameters is matching reasoning systems hundreds of times larger on hard math. VibeThinker-3B, built by Sina Weibo and released under the permissive MIT license, scores 94.3 on the AIME26 math benchmark — edging out DeepSeek V3.2 (671 billion parameters) at 94.2 and Kimi K2.5 (1 trillion) at 93.3. On a held-out set of unseen LeetCode contests, it passed 123 of 128 problems on the first attempt.

The size gap is the headline. VibeThinker isn't trained from scratch; it's a post-trained specialist built on Qwen2.5-Coder-3B. Its BF16 weights weigh in around 6 GB, small enough to run on a single consumer GPU. A model you can fit on one card is competing, on a specific kind of problem, with models that need a server rack.

How a small model punches up

The trick is the training recipe, which the team calls Spectrum-to-Signal. In plain terms: first teach the model a wide spread of possible solution paths instead of one rote answer (the "spectrum"), then use reinforcement learning to sharpen the ones that actually work (the "signal"), focusing effort on problems right at the edge of what the model can currently do. There's also an optional test-time step that generates several attempts, checks its own intermediate claims, and votes on the most reliable answer — pushing AIME26 as high as 97.1 with no extra parameters.

It's worth being precise about where it shines and where it doesn't. On math and competitive coding, VibeThinker-3B trades blows with the giants. On broad knowledge tests like GPQA-Diamond it scores 70.2, well behind DeepSeek's 82.4 and Kimi's 87.6 — because raw factual recall really does scale with size, and 3 billion parameters can only hold so much. This is a sharp tool, not a Swiss Army knife.

Why it matters for you

The practical takeaway is about cost and control. If a 3B model can handle your narrow task — math, code, structured reasoning — you can run it locally, privately, and for roughly the price of electricity, instead of paying per token to a frontier API. For anyone building a product on top of AI, that's the difference between a feature that's free to run and one that bleeds money at scale.

It also chips away at the assumption that bigger is always better. The story of 2026 isn't just trillion-parameter labs racing each other; it's small, specialized models getting good enough that the frontier stops being the only option. MIT licensing means anyone can take VibeThinker, inspect it, fine-tune it, and ship it — no permission, no usage cap, no kill switch.

My read: don't reach for the biggest model out of habit. For a tightly scoped job, a tuned small model is often faster, cheaper, and entirely yours — and 'entirely yours' is worth a lot when an API you don't control can change its price or pull a model overnight.