ML research
VibeThinker-3B small reasoning model
- Category: ML research
- Status: developing
- Sources: arXiv 2606.16140, discussion
- Summary: A team led by Sen Xu describes VibeThinker-3B, a 3-billion-parameter reasoning model trained with a "Spectrum-to-Signal" post-training recipe: curriculum-based supervised fine-tuning, multi-domain reinforcement learning, then offline self-distillation. The paper (submitted 2026-06-15) reports AIME26 94.3 (97.1 with test-time scaling), LiveCodeBench v6 80.2 Pass@1, LeetCode-contest 96.1% acceptance, and IFEval 93.4, and claims parity on these tasks with models "orders of magnitude larger" such as DeepSeek V3.2, GLM-5, and Gemini 3 Pro. The reported numbers are the authors' own and not independently reproduced.
- Why it matters: If the recipe reproduces, it pushes competition-grade math and coding reasoning into a 3B model that runs on commodity hardware, lowering the cost floor for local reasoning workloads.
- Follow-up: Watch for released weights, the license, and independent evaluation on held-out contest problems.