Best local LLM for reasoning
Last updated 2026-05-26 · Page updated 2026-07-13
Top 6 open-source picks for reasoning and math, ranked by benchmark performance and real-world fit. Updated monthly.
DeepSeek R1 Distill 32B
The 32B DeepSeek R1 distill — the best accessible open-weight reasoner we've tested. Explicit chain-of-thought, MIT-licensed, runs on a single 24GB GPU.
QwQ 32B
Alibaba's dedicated 32B reasoner, trained with reinforcement learning rather than distillation. Hits 79.5 on AIME24 and 90.6 on MATH-500 — a direct Apache-licensed alternative to DeepSeek R1.
Phi-4 Reasoning 14B
Microsoft's 14B reasoner that beats R1-Distill-Llama-70B on AIME and GPQA with 50x fewer parameters. MIT-licensed, English-first, with a 32K context.
DeepSeek R1 Distill Qwen 14B
DeepSeek's R1 reasoning distilled into Qwen 14B under MIT. AIME24 69.7 and MATH-500 93.9 — beats o1-mini on most reasoning benchmarks.
DeepSeek R1 Distill 7B
A 7B DeepSeek model distilled from R1 671B with explicit chain-of-thought reasoning. Surprisingly strong on AIME and MATH for its size.
Qwen 3 30B-A3B
Alibaba's Qwen 3 MoE with 30B total and just 3B active parameters, supporting hybrid thinking mode. MMLU 81.4, AIME24 80.4, 100+ languages, Apache 2.0.
Which GPU should you buy to run DeepSeek R1 Distill 32B?
To run DeepSeek R1 Distill 32B locally at Q4, you need ~19 GB of VRAM. The best value for this is a RTX 4090 (24 GB VRAM).
As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.
Frequently asked questions
What is the best local LLM for reasoning and math?
DeepSeek R1 Distill 32B tops this ranking — a 32B model, licensed under MIT, needing about 19 GB of VRAM at Q4 quantization. See the full list below for the runner-ups and how they compare.
How much VRAM do I need to run DeepSeek R1 Distill 32B?
At Q4 quantization, DeepSeek R1 Distill 32B needs about 19 GB of VRAM and fits comfortably on a single 24 GB GPU.
Which of these models fit an 8 GB GPU?
At Q4 quantization, DeepSeek R1 Distill 7B fit within 8 GB of VRAM.
Are the models on this reasoning and math list free for commercial use?
Licenses across this list include Apache 2.0, MIT. Check the specific license of each model on its catalog page before deploying commercially, as terms vary by author.
What context window do these models support?
Context windows on this list range from 32k to 128k tokens, depending on the model.