VibeThinker-3B is a 3-billion-parameter reasoning-focused language model developed by WeiboAI, built on top of Qwen2.5-Coder-3B and trained using the Spectrum-to-Signal Principle (SSP) post-training pipeline, which combines curriculum-based two-stage SFT, multi-domain reinforcement learning via MaxEnt-Guided Policy Optimization (MGPO), offline self-distillation, and instruct RL to systematically develop strong verifiable reasoning capabilities across mathematics, coding, and STEM tasks. Motivated by the Parametric Compression-Coverage Hypothesis — which posits that verifiable reasoning is a highly compressible, parameter-dense capability that compact models can carry near-frontier performance in — VibeThinker-3B achieves remarkable results for its size, scoring 76.4 on IMO-AnswerBench (improving to 80.6 with Claim-Level Reliability Assessment test-time scaling), competing with models like DeepSeek V3.2 (671B) and Kimi K2.5 (1T), while also achieving a 96.1% acceptance rate (123/128 submissions) on recent unseen LeetCode weekly and biweekly contests from April–May 2026, and reaching the performance range of top-tier frontier reasoning systems including Qwen3.6 Plus and Gemini 3 Pro on verifiable reasoning benchmarks. Inference is recommended via vLLM or SGLang with temperature 1.0, top_p 0.95, and supports up to 102K output tokens, making it best suited for competitive mathematics, coding contests, and STEM reasoning tasks where clear verification signals exist, while larger general-purpose models remain preferable for broad open-domain knowledge tasks.
