Learn About Weibo's AI Products
Create a free account to access in-depth lessons on each tool and model.
Start Learning FreeπAbout Weibo
Updated June 23, 2026Weibo is one of China's largest social-media platforms β a microblogging service often compared to X (formerly Twitter) β founded in 2009 and operated by Sina Weibo Inc. The company is publicly traded on both the Nasdaq (ticker WB) and the Hong Kong Stock Exchange, and counts Alibaba among its major shareholders. While Weibo's core business is social media and digital advertising, it has become an unexpectedly notable name in open AI research through its in-house large language model team, WeiboAI.
WeiboAI is best known for the VibeThinker series of small, open-weight reasoning models. In November 2025 it released VibeThinker-1.5B, a 1.5 billion-parameter model that matched or beat the original DeepSeek R1 on several math benchmarks for a reported post-training cost of about $7,800 β a tiny fraction of what comparable models cost. In June 2026 the team followed with VibeThinker-3B, a 3.1 billion-parameter model built on Alibaba's Qwen2.5-Coder-3B base and released under the permissive MIT license, posting frontier-level math and coding-reasoning scores for its size.
Both models lean on a training recipe the team calls the Spectrum-to-Signal Principle β curriculum fine-tuning followed by reinforcement learning β to coax strong step-by-step reasoning out of tiny models that can run on a single consumer GPU. The work has drawn attention precisely because it challenges the assumption that frontier reasoning requires enormous scale and budgets. It has also drawn scrutiny, with some researchers questioning whether the headline benchmark numbers hold up outside the narrow tasks the models were tuned on. The research team is small β fewer than ten authors, all Sina Weibo employees β which makes the results all the more striking for a company whose main product is a social feed, not a frontier lab.
π οΈProducts & Tools (1)
Weibo AI's 3.1B-parameter open reasoning model (MIT, built on Qwen2.5-Coder-3B). Runs on a single consumer GPU; posts strong but contested math and coding-reasoning benchmark scores.
