You do not need a data centre to run a capable AI model. Open-weight models run on a gaming PC or a recent Mac, for free and without sending anything online. The real question is which ones fit. We took the 61 open models that have an independent score, worked out how much memory each needs once compressed to 4 bits (the usual way to run them at home), and matched them against common amounts of memory.
What fits on each graphics card
| Memory | Models that fit | Best of them (overall index) |
|---|---|---|
| 8 GB graphics card | 7 | Qwen3.5-9B (139.5) |
| 12 GB graphics card | 11 | Qwen3.5-9B (139.5) |
| 16 GB graphics card | 12 | Qwen3.5-9B (139.5) |
| 24 GB graphics card | 23 | Qwen3.8 27B (149.4) |
| 32 GB graphics card | 25 | Qwen3.8 27B (149.4) |
| Mac with 32 GB | 23 | Qwen3.8 27B (149.4) |
| Mac with 64 GB | 27 | Qwen3.8 27B (149.4) |
| Mac with 128 GB | 31 | Qwen3.8 27B (149.4) |
Even 8 GB is useful: 7 rated models fit, and the best of them, Qwen3.5-9B, scores 139.5 on our overall index. The step that matters most is 24 GB: 23 models fit, including Qwen3.8 27B (149.4). Beyond that, extra memory buys surprisingly little among rated models, because the next tier of open models is far larger: Kimi K3, the best open model, needs server-class hardware.
How close is home AI to the best?
The best model that fits on a 24 GB card scores 149.4. The best open model overall scores 157.7, and the best model we track, GPT-6 Astra, scores 166.6. So a strong home setup gets within 17.2 points of the frontier. For summaries, drafts, translation and questions on your own documents, that is usually enough. For hard reasoning and long pieces of code, the large cloud models still lead.
What these numbers assume
- 4-bit compression, about 0.6 bytes per parameter, plus about 1.5 GB for an ordinary conversation. Longer documents need more.
- For a model to be fast, it has to fit in the graphics card’s memory. If it does not, it can spill into the computer’s main memory, but it then writes slowly.
- On a Mac with an Apple chip, about 70% of the unified memory can go to the graphics side.
- Only models with an independent score are ranked here. Many smaller or newer models exist too: the tool lists all 148 open models whose size we know.
Check your own machine with the Can my PC run it? tool, and follow the guide on how to run a model on your computer.
The figures in this article are recalculated from our data at every update (last: Sep 30, 2026). How we work