Best local LLM: what people run, by memory size
The best local LLM is the biggest one that fits your memory. Below is the model people report most for each memory size, with the speed they get, and a ranking across every setup.
Measured by readers on their own hardware. Newest report: 2026-11-11.
The pick for each memory size
The most-reported model that fits whole in that much graphics or unified memory. Speed is the median of plain runs.
Most reported across every setup
One row per model size. Speculative runs are counted apart from plain ones.
| # | Model | Reports | Gen t/s | PP t/s | With speculation |
|---|---|---|---|---|---|
| 1 | Qwen3.827B Alibaba · mostly Q4_K_M Gen 27 t/s (140 runs) · PP 433 t/s (65 runs) · with speculation 67 t/s (185) | 661 | 27 140 runs, 52 devices | 433 65 runs | 67 185 runs |
| 2 | Qwen3.8 Flash-Next125B · 6B active Alibaba · mostly UD-Q4_K_XL Gen 40 t/s (63 runs) · PP 1200 t/s (41 runs) · with speculation 47 t/s (62) | 574 | 40 63 runs, 13 devices | 1200 41 runs | 47 62 runs |
| 3 | Qwen3.635B · 3B active Alibaba · mostly Q4_K_M Gen 61 t/s (68 runs) · PP 1013 t/s (44 runs) · with speculation 121 t/s (21) | 205 | 61 68 runs, 23 devices | 1013 44 runs | 121 21 runs |
| 4 | DeepSeek V4 Flash284B · 13B active DeepSeek · mostly MXFP4 Gen 29 t/s (18 runs) · PP 312 t/s (10 runs) · with speculation 29 t/s (7) | 133 | 29 18 runs, 9 devices | 312 10 runs | 29 7 runs |
| 5 | Qwen3.627B Alibaba · mostly Q4_K_M Gen 26 t/s (20 runs) · PP 341 t/s (14 runs) · with speculation 50 t/s (19) | 77 | 26 20 runs, 13 devices | 341 14 runs | 50 19 runs |
| 6 | DeepSeek V4.1 Flash552B · 16B active mostly IQ2_XXS Gen 22 t/s (6 runs) · PP 1119 t/s (4 runs) · with speculation 31 t/s (3) | 76 | 22 6 runs, 4 devices | 1119 4 runs | 31 3 runs |
| 7 | Gemma 426B · 4B active Google DeepMind · mostly Q4_K_M Gen 79 t/s (33 runs) · PP 2595 t/s (25 runs) · with speculation 340 t/s (2) | 56 | 79 33 runs, 17 devices | 2595 25 runs | 340 2 runs |
| 8 | GLM-5.3 Flash320B · 18B active Zhipu AI · mostly NVFP4 Gen 37 t/s (5 runs) · PP 875 t/s (2 runs) · with speculation 70 t/s (4) | 51 | 37 5 runs, 3 devices | 875 2 runs | 70 4 runs |
| 9 | Qwen3.59B Alibaba · mostly Q4_K_M Gen 48 t/s (19 runs) · PP 826 t/s (8 runs) · with speculation 116 t/s (3) | 46 | 48 19 runs, 14 devices | 826 8 runs | 116 3 runs |
| 10 | Qwen3.8 Swift-1.527B Alibaba · mostly IQ3_S Gen 45 t/s (4 runs) · with speculation 83 t/s (13) | 36 | 45 4 runs, 4 devices | no PP data | 83 13 runs |
| 11 | Gemma 431B Google DeepMind · mostly Q4_K_M Gen 13 t/s (16 runs) · PP 516 t/s (12 runs) · with speculation 38 t/s (4) | 30 | 13 16 runs, 11 devices | 516 12 runs | 38 4 runs |
| 12 | Qwen3.535B · 3B active Alibaba · mostly Q4_K_M Gen 60 t/s (17 runs) · PP 1770 t/s (13 runs) | 27 | 60 17 runs, 12 devices | 1770 13 runs | none |
| 13 | Gemma 4 E4B8B Google DeepMind · mostly Q4_K_M Gen 51 t/s (12 runs) · PP 787 t/s (7 runs) | 26 | 51 12 runs, 9 devices | 787 7 runs | none |
| 14 | Qwen3.54B Alibaba · mostly Q4_K_M Gen 85 t/s (10 runs) · PP 1522 t/s (6 runs) · with speculation 11 t/s (1) | 23 | 85 10 runs, 8 devices | 1522 6 runs | 11 1 run |
| 15 | Qwen3-Coder30B · 3B active mostly UD-Q4_K_XL Gen 69 t/s (15 runs) · PP 578 t/s (11 runs) | 23 | 69 15 runs, 5 devices | 578 11 runs | none |
How we rank
Models are ordered by how many reports name them, because people report what they actually run. The speed beside each one is the median of plain runs only. Runs that used speculative decoding, several requests at once or CPU offload are counted apart. A model counts as fitting a memory size when the calculator places it whole in that memory. We don't score answer quality, so a popular model is not a claim that it is the best writer.
Frequently asked
What is the best local LLM?
There is no single answer, because the model that fits a 24GB card does not fit an 8GB one. Find your memory size in the table above and read the pick for it. The picks rank how many people report a model, so they show what people use, not which model writes best.
How do you rank them?
By how many reports name each model size, with the median speed of plain runs beside it. A run that used speculative decoding, several requests at once or CPU offload is counted apart and never mixed into the median.
Is the most-reported model the best one?
No. A model people report often is popular and easy to run. Quality is a separate question, and we make no claim about it. Try a model on a task of your own before you decide.
What is the best local LLM for coding?
It also depends on your memory. The coding page splits what people run for code by memory size.