A 3-billion-parameter model has outperformed Google's Gemini 3 Pro on the AIME 2026 math competition, achieving a score of 94.3 compared to Gemini 3 Pro's 91.7. This smaller model, developed by Weibo and released under an MIT license, is surprisingly effective at complex mathematical reasoning, challenging expectations about the capabilities of models with fewer parameters. AI
IMPACT Demonstrates that smaller, potentially more efficient models can achieve high performance on complex reasoning tasks, challenging the trend towards ever-larger models.
RANK_REASON A research paper details a model's performance on a benchmark, comparing it to a known model. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
