VibeThinker-3B shows small-model coding power thanks to aggressive post-training
VibeThinker-3B, a 3.09 B‑parameter model built on Qwen2.5‑Coder, narrows the gap to much larger coding‑reasoning systems by leveraging curated synthetic math/code data and a two‑stage finetuning pipeline with aggressive filtering and RL tricks. The effort cost roughly $25‑60 k, making high performance affordable, though real‑world tests are still pending.