Sina Weibo’s 3B-Parameter Model Outperforms Giants on AI Benchmarks, Sparks Debate

A 3-billion-parameter AI model developed by Sina Weibo, the Chinese social media giant, has achieved scores on math and coding benchmarks that rival or surpass models from Google DeepMind, OpenAI, and others with hundreds of times more parameters, according to a technical report published on arXiv. The model, called VibeThinker-3B, scored 94.3 on the 2026 American Invitational Mathematics Examination (AIME), outperforming DeepSeek V3.2 (91.7) and matching Gemini 3 Pro, according to the paper.

How Did a 3B-Parameter Model Outperform Giants?

VibeThinker-3B’s performance defies conventional scaling laws, which suggest larger models generally outperform smaller ones. The model achieved 91.4 on AIME 2025, 93.8 on the Brown University Math Olympiad, and an 80.2 Pass@1 score on LiveCodeBench v6, a coding benchmark. Its 96.1% success rate on LeetCode contests from April to May 2026 further highlights its capabilities, according to the report.

The model is based on Alibaba’s Qwen2.5-Coder-3B and refined through a four-stage training pipeline, including reinforcement learning and distillation of high-quality reasoning trajectories. Researchers at Sina Weibo describe this as evidence for the “Parametric Compression-Coverage Hypothesis,” which posits that reasoning tasks can be compressed into smaller models, while knowledge-intensive tasks require larger systems.

Community Skepticism: Are Benchmarks Reliable?

Despite the impressive scores, the AI community has expressed skepticism. Critics argue that benchmarks like AIME and LeetCode may be “gameable,” with models optimized for specific patterns rather than real-world utility. “The benchmarks are literal pattern matching for single-file coding. It has no relation to actual coding work,” wrote user @BigMoonKR on X, which has over 161,000 views.

Some users who tested the model reported practical shortcomings. One noted the model “doesn’t even know what a UV script is,” a popular Python tool. Others questioned the validity of the LeetCode results, suggesting they might reflect “benchmark leakage” rather than genuine capability. The paper claims training sets were decontaminated using n-gram filtering, but real-world utility remains unproven.

Implications for AI Development: Smaller Models Could Reshape the Industry

VibeThinker-3B’s success challenges the dominance of large-scale models, which require massive computational resources. Sina Weibo’s team emphasized that the model’s achievements are task-specific, noting it scored 70.2 on GPQA-Diamond, a science knowledge benchmark, compared to 91.9 for Gemini 3 Pro. “The true significance lies in showing compact models can excel on tasks with clear verification signals,” the paper states.

Experts like Francesco Bertolotti, an AI researcher, highlighted the engineering feat: “Even if it’s benchmaxed, doing this with 3B parameters is fascinating.” The model’s open-source release under the MIT License and free availability on platforms like Hugging Face have sparked interest in hybrid AI architectures, where small, specialized models handle reasoning tasks while larger systems provide factual knowledge.

What’s Next for AI Research?

The debate over model size versus efficiency is intensifying as companies race to balance performance with cost. VibeThinker-3B’s creators argue that its approach could reduce deployment costs, making advanced reasoning capabilities accessible on consumer hardware. However, the AI industry remains divided on whether such models can replace large-scale systems in real-world applications.

As the field evolves, the question of whether “bigger is better” may soon be answered by the practicality of models like VibeThinker-3B. For now, the paper has forced a critical reevaluation of how AI progress is measured—and what it truly means for the future of the industry.