Best Local LLM to Run AI on Your Hardware (2026)

Best Local Models at a Glance

The short answer: GPT-oss 20B leads local coding models at 98.3% on the 38-task benchmark and needs only about 12GB of RAM, which makes it the best all-around pick for 2026. Prefer Qwen? Qwen 3.5 35B scores 85.8% but wants 22GB. On a low-RAM laptop, Gemma 3 12B still hits 80.6% in roughly 3GB. Pick by the RAM you actually have, then use the comparison below for quant options and speed.

  1. GPT-oss 20B: 98.3% (38-task benchmark), ~12 GB RAM
  2. Qwen 3.5 35B: 85.8% (38-task benchmark), 22 GB RAM
  3. Gemma 3 12B: 80.6% (38-task benchmark), ~3 GB RAM
  4. Qwen 3.5 27B: 75.2% (Aider edit benchmark), ~9 GB RAM
  5. DeepSeek Coder V2: 72.9% (Aider edit benchmark), ~6 GB RAM

Full methodology and per-task results are in the 38-task LLM benchmark. Weighing local against cloud models? The LLM picker tool ranks both by use case and budget.

Local LLM Comparison for Coding (2026)

Run these models on your own hardware. Scores from our 38-task benchmark and the Aider edit leaderboard.

ModelScoreSourceMin RAMQuant OptionsEst. Speed (32GB Apple Silicon)Install
GPT-oss 20B98.3%38-task benchmark11.7 GBQ2_K, Q3_K_S, Q3_K_M, Q4_0, IQ2_M, IQ2_S, IQ2_XS, IQ2_XXS, IQ3_M, IQ3_XS, IQ3_XXS, IQ4_NL, IQ4_XS, Q3_K_L, Q4_1, Q4_K_M, Q4_K_S, Q5_K_M, Q5_K_S, Q6_K, Q8_0, F16, BF16N/Aollama pull openai/gpt-oss-20b
Qwen 3.5 35B85.8%38-task benchmark22 GBQ4_K_M, Q8_0~17 tok/sollama pull qwen-local
Gemma 3 12B80.6%38-task benchmark3.1 GBIQ1_S, IQ1_M, IQ2_XXS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0, BF16~40 tok/sollama pull google/gemma-3-12b
Qwen 3.5 27B75.2%Aider edit benchmark8.8 GBIQ2_XXS, IQ2_M, IQ3_XXS, Q3_K_S, Q3_K_M, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0N/Aollama pull qwen3.5-27b
DeepSeek Coder V272.9%Aider edit benchmark6.1 GBIQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, IQ3_XS, IQ3_M, Q3_K_S, Q3_K_M, Q3_K_L, IQ4_XS, Q4_K_S, Q4_K_M, Q5_K_S, Q5_K_M, Q6_K, Q8_0, F32N/Aollama pull aider/deepseek-coder-v2
Qwen 3 32B72.9%Aider edit benchmark7.9 GBIQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0~17 tok/sollama pull qwen3-32b
Qwen 3 235B72.2%Aider edit benchmark140 GBQ4_K_MN/Aollama pull qwen3-235b
Qwen 2.5 Coder 32B71.4%Aider edit benchmark9.2 GBIQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0_4_4, Q4_0_4_8, Q4_0_8_8, Q4_0, Q4_K_S, Q4_K_M, Q5_K_S, Q5_K_M, Q6_K, Q8_0~17 tok/sollama pull aider/qwen2.5-coder-32b
Qwen 2.5 Coder 14B69.2%Aider edit benchmark4.8 GBIQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, Q4_0_4_4, Q4_0_4_8, Q4_0_8_8, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q5_K_S, Q5_K_M, Q6_K, Q8_0, F16N/Aollama pull aider/qwen2.5-coder-14b
DeepSeek R1 Distill 70B67.7%Aider edit benchmark16.3 GBIQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_MN/Aollama pull deepseek-r1-distill-70b
Qwen 3 30B-A3B65.4%Aider edit benchmark9.3 GBIQ1_S, IQ2_S, IQ2_XS, IQ1_M, IQ2_XXS, IQ2_M, Q2_K, IQ3_XS, IQ3_XXS, Q3_K_S, IQ3_M, Q3_K_L, Q3_K_M, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0N/Aollama pull qwen3-30b-a3b
DeepSeek R1 Distill 32B65.4%Aider edit benchmark9.2 GBIQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0~17 tok/sollama pull deepseek-r1-distill-32b
Qwen 3.5 14B63.9%Aider edit benchmark10 GBQ4_K_MN/Aollama pull qwen3.5-14b
Mistral Small 3.1 24B60.2%Aider edit benchmark5.7 GBIQ1_S, IQ1_M, IQ2_XXS, IQ2_M, Q2_K, IQ3_XXS, Q3_K_S, Q3_K_M, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0, BF16N/Aollama pull mistral-small-3.1-24b
Llama 3.3 70B59.4%Aider edit benchmark16.2 GBIQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, Q4_0_4_4, Q4_0_4_8, Q4_0_8_8, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_MN/Aollama pull aider/llama-3.3-70b
Llama 3.1 70B58.6%Aider edit benchmark17.2 GBIQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, Q4_K_S, Q4_K_M, Q5_K_SN/Aollama pull aider/llama-3.1-70b
Qwen 3 14B58.6%Aider edit benchmark3.9 GBIQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0, BF16N/Aollama pull qwen3-14b
Qwen 2.5 Coder 7B57.9%Aider edit benchmark2.8 GBIQ2_M, Q2_K, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, Q4_0_4_4, Q4_0_4_8, Q4_0_8_8, Q4_0, Q4_K_S, Q4_K_M, Q5_K_S, Q5_K_M, Q6_K, Q8_0, F16~60 tok/sollama pull aider/qwen2.5-coder-7b
Gemma 3 27B57.9%Aider edit benchmark6.7 GBIQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0N/Aollama pull gemma-3-27b
Phi-4 14B53.4%Aider edit benchmark5.7 GBQ2_K, Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0, F16N/Aollama pull phi-4-14b
Codestral 22B51.1%Aider edit benchmark6.8 GBIQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, Q4_K_S, Q4_K_M, Q5_K_S, Q5_K_M, Q6_K, Q8_0, F32~25 tok/sollama pull aider/codestral-22b
DeepSeek R1 Distill 14B49.6%Aider edit benchmark4.8 GBIQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0, F16N/Aollama pull deepseek-r1-distill-14b
Qwen 3 8B48.1%Aider edit benchmark2.3 GBIQ1_S, IQ1_M, IQ2_XXS, IQ2_M, Q2_K, IQ3_XXS, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_K_S, Q4_0, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0, BF16N/Aollama pull qwen3-8b
Llama 3.1 8B37.6%Aider edit benchmark3.0 GBIQ2_M, Q2_K, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0_4_4, Q4_0_4_8, Q4_0_8_8, Q4_K_S, Q4_K_M, Q5_K_S, Q5_K_M, Q6_K, Q8_0, F32N/Aollama pull aider/llama-3.1-8b
DeepSeek R1 Distill 7B36.1%Aider edit benchmark2.8 GBIQ2_M, Q2_K, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_1, Q5_K_S, Q5_K_M, Q6_K, Q8_0, F16, F32~60 tok/sollama pull deepseek-r1-distill-7b