3B Small Model's Programming Scores Rival Opus 4.5, Mysterious Model Sparks Heated Discussion, Turns Out to be Domestic
A 3B parameter dense reasoning model named VibeThinker-3B has gained significant attention for achieving performance comparable to leading models like Gemini 3 Pro, GPT-5 high, and Claude Opus 4.5 in verifiable reasoning tasks such as programming, mathematics, and STEM problem-solving, despite its significantly smaller size.
Developed by Sina Weibo's team, the model is built upon Qwen2.5-Coder-3B. Its training employs an upgraded Spectrum-to-Signal pipeline, featuring a curriculum-based two-stage supervised fine-tuning (SFT), multi-domain reinforcement learning (RL) inspired by MGPO, offline self-distillation, and instruction RL to enhance controllability. A key innovation is the Claim-Level Reliability (CLR) assessment, a test-time scaling strategy that further boosts performance on math benchmarks.
The model excels in specific, verifiable domains, scoring highly on tests like AIME26 (94.3/97.1 with CLR) and LiveCodeBench v6 (80.2 Pass@1). However, it performs less impressively in areas requiring broad general knowledge. The authors propose a "parameter compression coverage hypothesis," suggesting that verifiable reasoning abilities—reliant on multi-step logic and feedback—are highly compressible, while open-domain knowledge depends more on large-scale parameters.
VibeThinker-3B demonstrates that small models, when specialized for tasks with clear verification signals, can reach frontier performance, offering a complementary research path to scaling model size. The model is publicly available for download and testing.
marsbit39 dk önce
