✦Dispatch · 21 June 2026
The frontier fits in three billion parameters.
A nine-author technical report, posted to arXiv on 15 June, makes a claim that would have read as a typo a year ago: a dense model with three billion parameters — small enough to run on a phone — scoring in the same band as flagship reasoning systems hundreds of times its size. The model is called VibeThinker-3B, and the number that matters most isn’t a benchmark. It’s the parameter count.
✦What the numbers say
Frontier scores at floor-level size.
VibeThinker-3B was built to answer one question: how far can verifiable reasoning — math, code, the kind of work where an answer is either right or wrong — be pushed inside a strictly small model. The answer is uncomfortably far. It reaches 94.3 on AIME26, climbing to 97.1 with claim-level test-time scaling, 80.2 Pass@1 on LiveCodeBench v6, and a 96.1% acceptance rate on recent unseen LeetCode contests — problems it could not have memorized. A 93.4 on IFEval confirms the obvious worry away: all that reasoning didn’t cost it the ability to simply follow instructions.
benchmark VibeThinker-3B band it lands in
──────────────────────────────────────────────────────────────────
AIME26 94.3 first-tier reasoning
+ test-time scaling 97.1 (claim-level)
LiveCodeBench v6 80.2 Pass@1 first-tier coding
LeetCode (unseen) 96.1% accept out-of-distribution
IFEval 93.4 instruction control
compared against: DeepSeek V3.2 · GLM-5 · Gemini 3 Pro
parameters: 3B dense · orders of magnitude smaller
✦How it got there
Spread the spectrum, then find the signal.
The method is a post-training recipe the authors call the Spectrum-to-Signal paradigm. The intuition: a small model doesn’t have the room to brute-force every path, so you first teach it to explore a wide spectrum of possible reasoning approaches, then sharpen that spread down to the right one. In practice that means three stages stacked in order — curriculum-based supervised fine-tuning that ramps up difficulty, multi-domain reinforcement learning across math and code, and offline self-distillation to compress the result back into the model.
Underneath it sits what they name the Parametric Compression-Coverage Hypothesis — the idea that a small model’s limited parameters are better spent on covering the breadth of a problem space than on packing in raw knowledge. It is a bet that for verifiable reasoning, how you train matters more than how big you are. The benchmarks say the bet paid off: VibeThinker-3B matches or exceeds DeepSeek V3.2, GLM-5, and Gemini 3 Pro on these tasks — models that dwarf it.
✦How we’re reading this
The floor just learned to reason.
For most of the last two years, the argument against on-device intelligence was simple: the good models are enormous, and enormous models live in someone else’s data center. Real reasoning — the multi-step kind — was the last thing the cloud could still hold over the device in your pocket. VibeThinker-3B is a three-billion-parameter rebuttal. At that size, the weights fit on a laptop today and a phone tomorrow, with room to spare.
We don’t think a single model settles anything — the frontier in the cloud keeps rising too. But the gap between what you can rent and what you can ownis closing from the bottom, fast. That’s the whole premise of the layer we’re building: intelligence that runs where your data already lives, answers when the network can’t, and belongs to you. A 3B model that reasons like a giant is the floor rising to meet it.
✦ ✦ ✦
Reasoning used to mean scale.
Now it fits in your hand.