I Ran the 3B Model That Beat Gemini 3 Pro at Olympiad Math — It Shouldn't Work

2026-06-19 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, quick

Summary

The VibeThinker-3B, a 3-billion-parameter model, recently achieved a score of 94.3 on AIME 2026, surpassing Gemini 3 Pro's 91.7. This MIT-licensed model, whose 1.5B predecessor cost only \$7,800 to post-train, demonstrates first-tier performance in verifiable reasoning tasks like competition math and competitive coding. Crucially, its compact size allows it to run on a laptop, distinguishing it from larger general knowledge models. Its release on June 15, 2026, sparked immediate skepticism within the AI community, which has grown wary of "benchmark theater" and questions the real-world applicability of such specific benchmarks, leading to accusations of "benchmaxxing."

Key takeaway

For machine learning engineers evaluating models for specific, verifiable reasoning tasks like competitive programming or advanced mathematics, you should consider exploring smaller, specialized models. VibeThinker-3B demonstrates that top-tier performance in these domains is achievable with 3-billion-parameter models, offering significant advantages in deployment cost and local execution over larger, general-purpose alternatives. This shifts the focus from raw parameter count to task-specific optimization.

Key insights

A small, cost-effective 3B model achieved top-tier verifiable reasoning performance, challenging assumptions about model size and capability.

Principles

• Specific reasoning benchmarks can reveal capabilities not evident in general knowledge tests.
• Smaller models can achieve competitive performance on specialized, verifiable tasks.
• Cost-effective post-training can yield high-performing, deployable models.

In practice

• Run 3B models locally for competitive math or coding tasks.
• Evaluate specialized models against specific, verifiable reasoning benchmarks.

Topics

• VibeThinker-3B
• Small Language Models
• Mathematical Reasoning
• Competitive Programming
• Benchmark Evaluation
• Model Efficiency

Best for: AI Engineer, Research Scientist, Entrepreneur, AI Scientist, Machine Learning Engineer, AI Student

Related on AIssential

• GLM-5.2: Only a Few Months Behind Commercial Models — GLM-5.2 nears commercial model performance, while VibeThinker-3B shows small models can excel in verifiable reasoning via targeted training.
• TAI #193: Gemini 3.1 Pro Takes the Benchmarks Crown, but Can it Catch Up in the Tools Race? — Raw model intelligence does not always equate to practical utility in real-world applications.
• WHAT? Qwen3.6-A3B Could Solve This? — Instruction tuning and reasoning trace analysis can significantly enhance small MoE model performance on complex tasks.
• ASyMOB: Algebraic Symbolic Mathematical Operations Benchmark — LLMs struggle with symbolic math perturbations, indicating pattern memorization, though frontier models show surprising robustness.
• Why Weibo’s tiny VibeThinker-3B has the AI world arguing over benchmarks again — Small models can achieve top-tier verifiable reasoning performance, decoupling it from broad factual knowledge.
• Evaluation of LLMs for Mathematical Formalization in Lean — LLMs like Gemini 3.1 Pro and Claude Opus 4.7 show high effectiveness in generating formal Lean 4 proofs, with some models offering cost efficiency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.