Local LLM Hardware Requirements Calculator
Model sizes and memory figures checked 1 November 2026.
Let the app read your hardware
Run the free AIblogly PC Check app on Windows, then press the button. The app sends your RAM, graphics card and free disk space to this page through our server, where they are kept for a few minutes and then deleted.
Your hardware
What you type here stays in your browser; it is worked out on this page and sent nowhere.
What it can run
Choose your RAM to see results.
How this is worked out
A model runs when its file fits in your computer's memory. That is the whole rule; the tool above applies it to the 29 open models on our run an LLM locally guide.
- Runs well: the model fits in your dedicated graphics card's memory (VRAM) with a tenth of it left spare, because the card also has to hold the conversation the model is working on. On a Mac with Apple Silicon, where the system and the graphics chip share one pool, three quarters of the memory is counted.
- Runs, but slowly: the model is larger than the graphics memory but fits once ordinary RAM is added, with 4 GB left for the system. The part in RAM runs on the processor. A computer with integrated graphics, or no dedicated graphics card, runs everything this way.
- Possible with AirLLM, very slowly: the model is larger than all of your memory. AirLLM loads it from disk one layer at a time, which works and is very slow. Our low-end hardware guide has measured examples.
Within each group the newest model family comes first, and inside a family its largest model that fits: a larger model of the same family is the more capable one. Where a model is on the LMArena board, its rank is shown beside it.
Treat the result as a guide. The figures are for 4-bit versions of each model, the size most people run. A longer conversation needs more memory than a short one, and a higher-precision version of the same model needs more again. Where a model's maker publishes no memory figure, the tool estimates it from the download size and marks it "estimated".
Every model on the list
| Model | LMArena rank | Memory needed | Download | Ollama command |
|---|---|---|---|---|
| Gemma 4 E2B | - | 2.9 GB | 4.6-7.5 GB | gemma4:e2b |
| Gemma 4 E4B | - | 4.5 GB | 6.6-9.5 GB | gemma4:e4b |
| Gemma 4 12B | - | 6.7 GB | 7.7-8.0 GB | gemma4:12b |
| Gemma 4 26B A4B | #102 | 14.4 GB | 16-19 GB | gemma4:26b |
| Gemma 4 31B | #79 | 17.5 GB | 19-20 GB | gemma4:31b |
| Qwen3.5 2B | - | 3.5 GB | 2.7-3.1 GB | qwen3.5:2b |
| Qwen3.5 4B | - | 5.5 GB | 3.3-4.0 GB | qwen3.5:4b |
| Qwen3.5 9B | - | 6.5 GB | 6.6-7.6 GB | qwen3.5:9b |
| Qwen3.5 27B | #158 | 17 GB | 17-20 GB | qwen3.5:27b |
| Qwen3.5 35B A3B | #175 | 22 GB | 22 GB | qwen3.5:35b |
| Qwen3.8 27B | #111 | 16-19 GB | 18 GB | qwen3.8:27b |
| gpt-oss-20b | #272 | 16 GB | 14 GB | gpt-oss:20b |
| Nemotron 3 Nano 30B A3B | #276 | about 24 GB | 18-24 GB | nemotron-3-nano:30b |
| Mistral Small 3.1 24B | #291 | about 18 GB (estimated) | 15 GB | mistral-small3.1 |
| Gemma 3 12B | #238 | about 9.8 GB (estimated) | 8.2 GB | gemma3:12b |
| Granite 4.1 8B | #288 | about 6.4 GB (estimated) | 5.3 GB | granite4.1:8b |
| Gemma 3 4B | #289 | about 4.1 GB (estimated) | 3.4 GB | gemma3:4b |
| Llama 3.2 3B | #374 | about 2.4 GB (estimated) | 2.0 GB | llama3.2:3b |
| Llama 3.2 1B | #403 | about 1.6 GB (estimated) | 1.3 GB | llama3.2:1b |
| Kimi K3 | #20 | 610 GB (1-bit) | 1,510 GB file | - |
| Kimi K2.6 | #61 | 350 GB (2-bit) | 600 GB | - |
| GLM-5.3 | #28 | 223 GB (1-bit) | 372-475 GB | - |
| DeepSeek V4.1-Flash | #38 | 128-145 GB (1-bit, Engram table on SSD) | 305 GB | - |
| GLM-5.3-Flash | #36 | 100 GB (1-bit) | 162-210 GB | - |
| MiniMax-M3 | #106 | 133 GB (1-bit) | 213-270 GB | - |
| Qwen3.5 397B A17B | #97 | 180 GB (3-bit) | 214 GB | - |
| DeepSeek V4-Flash | #105 | 92 GB (1-bit) | 162 GB | - |
| Qwen3.5 122B A10B | #140 | 60 GB (3-bit) | 70 GB | - |
| gpt-oss-120b | #220 | 66 GB (for 6+ tok/s) | 65 GB download (Ollama) | - |
Sources for every figure are on How to run an LLM locally. For the largest models the memory shown is for the smallest published version.
Install the one you picked
- How to install Ollama: one command to install, one to run a model.
- How to install LM Studio: a desktop app with a model browser built in.
- How to install llama.cpp: the engine underneath, for full control.
- How to install AirLLM: for models larger than your memory.
- Open-source LLM leaderboard: how these models rank against each other.
Let the free app read your hardware
AIblogly PC Check is a free Windows app that reads your RAM, graphics card, graphics memory and free disk space. Run it, leave its window open, and press Check my PC at the top of this page: a progress bar follows the app and the form fills itself in.
Download AIblogly PC Check for Windows (version 0.2.1, 31 KB, Windows 10 and 11, nothing to install)
- What it reads: Windows version, processor, RAM, graphics cards and their memory, free disk space, and whether Ollama, LM Studio and Python are already installed.
- Windows will warn you the first time. The app is not yet signed with a publisher certificate, so Windows shows "Windows protected your PC". Choose More info, then Run anyway.
- File check (SHA-256):
78541952dd57b07beef48c02cb9cd944fc2868d55b59480fcfe47d3caaffc03f
Installing a model from inside the app, with Ollama, LM Studio or AirLLM, is coming soon. On a Mac or Linux, enter your hardware in the form above.
Frequently asked questions
What LLM can I run with 8 GB of VRAM?+
Models that need less than 8 GB fit entirely on the card. On our list that includes Gemma 4 12B, which Google lists at 6.7 GB in 4-bit, and Qwen3.5 9B, which Unsloth lists at 6.5 GB. Larger models still run by using ordinary RAM as well, more slowly.
Can I run an LLM with 16 GB of RAM and no graphics card?+
Yes, on the processor, more slowly than on a graphics card. Leaving about 4 GB for the system, models that need up to about 12 GB fit, such as Gemma 4 12B at 6.7 GB.
How much memory does a Mac need to run an LLM?+
A Mac with Apple Silicon shares one pool of memory between the system and the graphics chip. This page counts three quarters of it as available to a model, so a 16 GB Mac is treated as having about 12 GB.
Is my hardware information sent anywhere?+
Not when you type it in: the result is worked out in your browser, and the figures appear only in the page address so you can bookmark or share a result. If you use the AIblogly PC Check app with the Check my PC button, the app sends your RAM, graphics card and free disk space to this page through our server, which keeps them for a few minutes and then deletes them.