VibeThinker-3B

VibeThinker-3B, released June 2026 by Sina Weibo's WeiboAI team, is a compact open-source reasoning model built on Qwen2.5-Coder as its base. It is reported to match models like DeepSeek V3.2 (671B), GLM-5 (744B), and Gemini 3 Pro on verifiable reasoning benchmarks — hundreds of times its size — from a post-training budget of roughly $7,800 for its smaller 1.5B sibling. The project is fully open: weights on Hugging Face and training code on GitHub.

The honest picture: the flagship-matching claims center on verifiable reasoning tasks, not broad general capability — a 3B model will not match a 671B one on open-ended generation or knowledge breadth. Independent replication of the headline results is still thin. But as a demonstration of what post-training can squeeze from small bases, it is one of the more interesting open releases of 2026.