- The best model we've tested is Qwen3.6 27B, with 97 out of 100.
- Close behind: gpt-oss-120b (96).
- Less memory? Find your graphics card or Mac below. We list the best model that fits and which version to download.
Pick by graphics card
Not sure how much memory your card has? On Windows, open Task Manager, go to Performance and click GPU. More help here.
Pick by Mac
Click the Apple menu, then About This Mac. The number next to Memory is what you need.
All models we've tested
| # | Model | Score | Coding | Documents | Hard tasks | Runs on (PC) | Runs on (Mac) |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3.6 27B 28B | 97 | 97 | 97 | 91 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 2 | gpt-oss-120b 117B · gpt-oss:120b | 96 | 93 | 99 | 94 | more than 32 GB | 96 GB Mac (standard) |
| 3 | Muse Glimmer 30B 30B | 94 | 90 | 98 | 89 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 4 | Ling 3.0 Flash 127B | 94 | 93 | 95 | 91 | more than 32 GB | 128 GB Mac (Q4) |
| 5 | GLM 4.5V 108B | 93 | 87 | 99 | 89 | more than 32 GB | 96 GB Mac (Q4) |
| 6 | Ling 3.0 Flash VL 125B | 92 | 87 | 98 | 83 | more than 32 GB | 128 GB Mac (Q5) |
| 7 | gpt-oss-20b 21B · gpt-oss:20b | 88 | 90 | 87 | 69 | 16 GB card (standard) | 24 GB Mac (standard) |
| 8 | Qwen3.5-27B 28B | 88 | 77 | 100 | 78 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 9 | Qwen3.8 27B 28B | 88 | 77 | 99 | 83 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 10 | Qwen3.5-35B-A3B 36B | 88 | 80 | 95 | 73 | 24 GB card (Q4) | 48 GB Mac (Q6) |
| 11 | Qwen3.6 35B A3B 36B | 86 | 80 | 93 | 73 | 24 GB card (Q4) | 48 GB Mac (Q6) |
| 12 | Laguna XS 2.1 33B | 82 | 73 | 91 | 56 | 24 GB card (Q4) | 48 GB Mac (Q6) |
| 13 | Nemotron 3 Nano 30B A3B 32B | 80 | 70 | 91 | 69 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 14 | Gemma 4 31B 31B | 79 | 93 | 65 | 73 | 32 GB card (Q5) | 48 GB Mac (Q6) |
| 15 | Granite 4.2 8B 8.8B | 77 | 77 | 78 | 55 | 8 GB card (Q4) | 16 GB Mac (Q6) |
| 16 | Qwen3 30B A3B 31B · qwen3:30b | 77 | 67 | 87 | 50 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 17 | Qwen3 VL 30B A3B Thinking 31B | 77 | 60 | 93 | 51 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 18 | GLM 4.5 Air 110B | 76 | 70 | 82 | 52 | more than 32 GB | 96 GB Mac (Q4) |
| 19 | GLM 4.6V 108B | 73 | 73 | 73 | 53 | more than 32 GB | 96 GB Mac (Q4) |
| 20 | Qwen3 14B 15B · qwen3:14b | 72 | 63 | 81 | 42 | 12 GB card (Q4) | 24 GB Mac (Q6) |
| 21 | GLM 4.7 Flash 31B | 72 | 73 | 70 | 43 | 24 GB card (Q4) | 48 GB Mac (Q6) |
| 22 | Qwen3.5-9B 9.7B | 72 | 47 | 97 | 58 | 8 GB card (Q4) | 16 GB Mac (Q6) |
| 23 | Gemma 4 26B A4B 26B | 70 | 83 | 58 | 50 | 24 GB card (Q5) | 32 GB Mac (Q5) |
| 24 | Qwen3 32B 33B · qwen3:32b | 64 | 50 | 79 | 50 | 24 GB card (Q4) | 48 GB Mac (Q6) |
| 25 | Laguna S 2.1 118B | 64 | 70 | 59 | 46 | more than 32 GB | 128 GB Mac (Q5) |
| 26 | Nemotron 3.5 Lightning 32B | 57 | 43 | 72 | 43 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 27 | Nex-N2.5-Mini 35B | 57 | 27 | 88 | 40 | 24 GB card (Q4) | 48 GB Mac (Q6) |
| 28 | Devstral 2 2512 125B | 53 | 53 | 53 | 24 | more than 32 GB | 128 GB Mac (Q5) |
| 29 | Hunyuan A13B Instruct 80B | 52 | 37 | 67 | 28 | more than 32 GB | 96 GB Mac (Q6) |
| 30 | Qwen3 Coder 30B A3B Instruct 31B · qwen3-coder:30b | 51 | 57 | 46 | 42 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 31 | Qwen3 VL 30B A3B Instruct 31B | 46 | 43 | 49 | 33 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 32 | Qwen2.5 72B Instruct 73B | 46 | 40 | 52 | 23 | more than 32 GB | 64 GB Mac (Q4) |
| 33 | Llama 3.3 70B Instruct 71B | 46 | 47 | 45 | 27 | more than 32 GB | 64 GB Mac (Q4) |
| 34 | Command A 111B | 45 | 33 | 56 | 27 | more than 32 GB | 96 GB Mac (Q4) |
| 35 | Ministral 3 14B 2512 14B | 44 | 43 | 45 | 24 | 12 GB card (Q4) | 16 GB Mac (Q4) |
| 36 | Mistral Small 3.2 24B 24B | 44 | 40 | 47 | 21 | 24 GB card (Q6) | 32 GB Mac (Q5) |
| 37 | Mistral Small 3 24B | 41 | 37 | 45 | 24 | 24 GB card (Q6) | 32 GB Mac (Q5) |
| 38 | Gemma 3 27B 27B · gemma3:27b | 40 | 33 | 46 | 29 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 39 | Qwen3 30B A3B Instruct 2507 31B | 40 | 33 | 46 | 25 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 40 | Qwen3 VL 8B Instruct 8.8B | 39 | 37 | 41 | 25 | 8 GB card (Q4) | 16 GB Mac (Q6) |
| 41 | Llama 3.1 70B Instruct 71B | 39 | 33 | 44 | 22 | more than 32 GB | 64 GB Mac (Q4) |
| 42 | Ministral 3 8B 2512 8.9B | 37 | 30 | 43 | 18 | 8 GB card (Q4) | 16 GB Mac (Q6) |
| 43 | Gemma 3 12B 12B · gemma3:12b | 35 | 30 | 40 | 24 | 12 GB card (Q5) | 16 GB Mac (Q5) |
| 44 | Phi 4 15B · phi4 | 25 | 23 | 26 | 19 | 12 GB card (Q4) | 24 GB Mac (Q6) |
| 45 | Qwen2.5 7B Instruct 7.6B · qwen2.5:7b | 24 | 10 | 37 | 11 | 8 GB card (Q6) | 16 GB Mac (Q8) |
| 46 | Gemma 2 27B 27B · gemma2:27b | 23 | 13 | 34 | 12 | 24 GB card (Q5) | 32 GB Mac (Q4) |
| 47 | Llama 3.3 8B Instruct (Q4, Mac) 8B · hf.co/bartowski/allura-forge_Llama-3.3-8B-Instruct-GGUF:Q4_K_M | 23 | 13 | 33 | 10 | 8 GB card (Q4) | 16 GB Mac (Q4) |
| 48 | Mistral Nemo 12B · mistral-nemo | 23 | 10 | 36 | 13 | 12 GB card (Q5) | 16 GB Mac (Q4) |
| 49 | Granite 4.0 Micro (Q4, Mac) 3.2B · granite4:micro-h | 22 | 13 | 31 | 14 | 8 GB card (Q4) | 16 GB Mac (Q4) |
| 50 | Ministral 3 3B 2512 3.9B | 21 | 10 | 32 | 15 | 8 GB card (Q8) | 16 GB Mac (Q8) |
| 51 | Llama 3.1 8B Instruct 8B · llama3.1:8b | 15 | 3 | 26 | 3 | 8 GB card (Q5) | 16 GB Mac (Q8) |
| 52 | Gemma 3 4B 4.3B · gemma3:4b | 14 | 3 | 25 | 6 | 8 GB card (Q8) | 16 GB Mac (Q8) |
| 53 | Qwen2.5 VL 72B Instruct 73B | 14 | 13 | 14 | 8 | more than 32 GB | 64 GB Mac (Q4) |
| 54 | Reka Edge 7.1B | 8 | 0 | 16 | 3 | 8 GB card (Q6) | 16 GB Mac (Q8) |
| 55 | Llama 3.2 3B Instruct 3.2B · llama3.2:3b | 6 | 0 | 13 | 2 | 8 GB card (Q8) | 16 GB Mac (Q8) |
| 56 | Llama 3.2 1B Instruct (Q4, Mac) 1.2B · llama3.2:1b-instruct-q4_K_M | 5 | 0 | 9 | 3 | 8 GB card (Q4) | 16 GB Mac (Q4) |
Scores are out of 100. "Runs on" is the smallest graphics card or Mac that can hold the model, and which version to download for it. How we worked this out
How we test
Every model gets the same 110 questions. 30 are small coding jobs, and we run the code to check that it works. The rest ask the model to read things like bills, pay stubs and email threads and pull out the right numbers. Many need a little math, or catching a correction further down the page.
We keep most questions secret, so no model can memorize the answers. The rest are public, along with every answer each model gave. Full technical details.
Questions people ask
What's the best local LLM for an 8 GB graphics card?
Right now it's Granite 4.2 8B. Download the Q4 version. It scored 77 out of 100 in our tests.
What's the best local LLM for a 12 GB graphics card?
Right now it's Granite 4.2 8B. Download the Q8 version. It scored 77 out of 100 in our tests.
What's the best local LLM for a 16 GB graphics card?
Right now it's gpt-oss-20b. Download the standard version. It scored 88 out of 100 in our tests.
What's the best local LLM for a 24 GB graphics card?
Right now it's Qwen3.6 27B. Download the Q5 version. It scored 97 out of 100 in our tests.
What's the best local LLM for a Mac with 16 GB of memory?
Right now it's Granite 4.2 8B. Download the Q6 version. It scored 77 out of 100 in our tests.
What's the best local LLM for a Mac with 32 GB of memory?
Right now it's Qwen3.6 27B. Download the Q4 version. It scored 97 out of 100 in our tests.
What do Q4 and Q8 mean?
They're versions of the same model, squeezed down to different sizes. Q4 is smaller and runs on more computers. Q8 is bigger and a little more accurate. If both fit on your computer, pick the bigger one.
Why don't you say how fast each model runs?
Speed depends a lot on your exact computer and the app you use, so any number we gave would be wrong for most people. We tell you which models fit on your machine and how good their answers are.
Comments
Sign in with GitHub to comment. Spam and abuse are hidden automatically.