The best local LLMs for math and reasoning tasks, ranked by VRAM tier. AIME and MATH benchmarks for DeepSeek R1, Qwen 3 thinking, and Phi-4-reasoning.