A little-known AI effort from Chinese social media company Sina Weibo has set off a fresh argument in the machine learning community about what benchmark scores actually measure. In a technical report posted to arXiv, a nine-researcher team said its 3 billion parameter model, VibeThinker-3B, can match or surpass much larger systems from Google DeepMind, OpenAI, Anthropic and DeepSeek on a range of reasoning tasks.
The eye-catching claim is not just the performance, but the size of the model. At 3 billion parameters, VibeThinker-3B is tiny by frontier-model standards and small enough to run on a consumer laptop. By comparison, some of the models it is being compared with have hundreds of billions of parameters.
The paper says VibeThinker-3B scored 94.3 on AIME 2026, a difficult mathematics benchmark. The researchers also reported strong results on other math sets, including AIME 2025, HMMT 2025, BruMO 2025 and IMO-AnswerBench. In coding, the model reached 80.2 Pass@1 on LiveCodeBench v6 and a 96.1 percent acceptance rate on recent LeetCode contests. It also scored 93.4 on IFEval for instruction following.
The results quickly circulated across AI forums and social media, where they were met with both curiosity and doubt. Some users questioned whether the benchmarks had become too easy to optimize for, especially in coding and math settings where the right answers can be checked automatically.
That skepticism reflects a broader concern in the AI field. Benchmark gains often create headlines, but researchers and users alike have grown wary of models that perform impressively on tests while struggling in practical use. Critics of VibeThinker-3B pointed to this gap, arguing that benchmark-heavy evaluations can miss important real-world capabilities such as tool use, long conversation coherence and everyday coding workflow support.
The paper’s authors say they tried to address contamination concerns by filtering training data for overlap with evaluation sets. They also point to the LeetCode contest results as especially meaningful because the contests were published in 2026, after any plausible training cutoff. According to the report, the model passed 123 of 128 first-attempt submissions in that setting.
The Weibo team frames its work around what it calls the Parametric Compression-Coverage Hypothesis. In simple terms, the paper argues that not every type of AI capability scales the same way with model size. The authors contend that verifiable reasoning, such as math and code problems with clear answers, can be compressed into smaller models more easily than open-ended knowledge tasks that depend on broad factual coverage.
That distinction matters because VibeThinker-3B did much better on math and coding than on general knowledge. On GPQA-Diamond, a graduate-level science benchmark, it scored 70.2, well behind much larger flagship models. The authors present that gap as evidence for their thesis, not a contradiction.
VibeThinker-3B is not a model built from scratch. It was post-trained from Alibaba’s Qwen2.5-Coder-3B using a multi-stage pipeline. The process included supervised fine-tuning, reinforcement learning across math, code and STEM tasks, distillation from reinforced checkpoints, and a final instruction-tuning stage.
The researchers say one key lesson was that methods that worked at the smaller 1.5B scale did not transfer cleanly to 3B. In particular, progressively expanding the context window during training hurt performance, so the team settled on a single 64,000-token context window throughout reinforcement learning.
The broader implication of the paper is clear: smaller models may be more capable than many observers assumed, at least on tasks where answers can be checked mechanically. But the online backlash also shows how far the field still is from agreeing on what counts as meaningful progress.
For now, VibeThinker-3B is less a settled breakthrough than a new flashpoint in an ongoing argument over whether AI benchmarks still tell the full story.