VibeThinker-3B beats models 300x larger — but only on math
sina's 3b vibecheck-3b matches deepseek v3.2 on aime26 math but fails gpqa knowledge benchmarks, proving logic compresses better than facts.
sina released vibecheck-3b, a 3b parameter model matching giants like deepseek v3.2 and kimi k2.5 on aime26 math and livecodebench despite being 333x smaller. built on alibaba's qwen2.5-coder-3b, its performance stems from multi-stage post-training combining supervised fine-tuning, reinforcement learning for math/code, and self-distillation. the model lags on gpqa-diamond, indicating that while logical reasoning compresses efficiently into small models, broad factual knowledge still demands large-scale parameters. researchers validated results with post-training leetcode contests where it solved 123 of 128 problems, outperforming gpt-5.2.