VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO

273 points · 128 comments on HN · read original →

Points and comments are a snapshot, not live.

A 3B-param model matches frontier reasoning on math and coding via SFT and GRPO.

VibeThinker-3B, a compact dense model, is trained with a post-training pipeline combining curriculum SFT, multi-domain RL, and offline self-distillation. It scores 94.3 on AIME26 (97.1 with test-time scaling), 80.2 on LiveCodeBench v6, and 96.1% on unseen LeetCode contests. This places it alongside models like DeepSeek V3.2, GLM-5, and Gemini 3 Pro. A 93.4 on IFEval shows reasoning gains don't impair instruction following. The Parametric Compression-Coverage Hypothesis posits that verifiable reasoning is compressible into compact cores, while broad knowledge requires more parameters.

What commenters are saying

Commenters emphasize the model is a specialist for closed-world, verifiable reasoning (math, coding contests) rather than general tasks like SVG generation. One user reports success using a quantized version for source code security review on an RTX 3090. Another tests it on a nasty ODE, finding a valid solution at 110 tok/s on a 2070 Super. Several note models like Qwen 3.5 9B also solve the same ODE. The thread distinguishes the model from general-purpose agents, suggesting it works best as a reasoning module in a larger system.