On this page
gpt-oss-20b runs on a Windows PC with 16 GB of RAM, with or without a graphics card. Off Grid AI Desktop runs it free and offline on Windows. Its 12 GB file sits entirely on a 16 GB NVIDIA or AMD card, and on smaller cards the app keeps part of it in system RAM. The 120B is a 63 GB file that needs about 96 GB of RAM.
You don’t need an OpenAI account, Python or WSL.
Download Off Grid AI Desktop (free, Windows x64)
What is gpt-oss, and which one fits your PC?
gpt-oss is OpenAI’s open-weight model family, released in August 2025 under Apache 2.0. There are two sizes (20B card, 120B card):
| gpt-oss-20b | gpt-oss-120b | |
|---|---|---|
| Parameters | 21B, 3.6B active per word | 117B, 5.1B active per word |
| File (MXFP4 GGUF) | 12.1 GB | 63.4 GB |
| OpenAI’s target | 16 GB of memory | One 80 GB GPU |
| Context | 131,072 tokens | 131,072 tokens |
| Reads images | No | No |
File sizes are from ggml-org’s 20B and 120B uploads, checked 16 November 2026.
Both are mixtures of experts. Only a few billion parameters work on each word, so a gpt-oss model copes far better than a dense model of the same size when part of it sits in system RAM.
What your GPU does with gpt-oss-20b
| Your GPU’s VRAM | Example cards | What happens |
|---|---|---|
| 24 GB or more | RTX 4090, 3090, 5090, 7900 XTX | Whole model on the card, room for long chats |
| 16 GB | RTX 5080, 5070 Ti, 5060 Ti 16 GB, 4060 Ti 16 GB | Whole 12.1 GB file on the card |
| 12 GB | RTX 5070, 4070, 3060 12 GB | Most of it on the card, some expert weights in RAM |
| 8 GB | RTX 4060, 3070, 5060 | Part on the card, the rest in system RAM |
| None | Any PC with 16 GB of RAM | Runs on the CPU |
With GPU layers on Auto, Off Grid AI Desktop lets the engine place the model, and for a mixture-of-experts model it can leave expert weights in system RAM when the card is short. You don’t set anything by hand.
System RAM still matters. Off Grid AI Desktop judges fit by file size against RAM: up to 65% runs well, up to 82% is Tight. The 12 GB file is Tight on a 16 GB PC and runs well from 24 GB.
How to set it up on Windows
- Download Off Grid AI Desktop and run the installer.
- On an NVIDIA card, go to Settings → GPU performance. In Use your NVIDIA GPU, choose Download (about 8 GB), then Restart app. AMD and Intel cards use Vulkan, which is built in.
- Open Models. gpt-oss isn’t in the app’s own list yet, so type gpt-oss-20b into Search models….
- Open the ggml-org or Unsloth upload and download a file of about 12 GB. Every gpt-oss-20b file is close to that size, because the expert weights already ship in 4-bit.
- Start a chat and pick the model. In Model settings → Text, leave the backend on Auto or choose CUDA on NVIDIA.
- Send a prompt, then check the backend’s Now value and GPU layers under Performance to see where it loaded.
A useful first prompt:
Write a PowerShell script that finds every file over 1 GB in my Downloads folder and lists them biggest first. Explain each line.
gpt-oss thinks before it answers. The app doesn’t offer a reasoning-effort switch for gpt-oss yet, so it runs at the model’s default, medium.
Once it’s downloaded, everything runs on your PC. Unplug the network and it keeps going.
What about gpt-oss-120b on Windows?
It runs on a workstation, not a typical gaming PC. The 63.4 GB file has to live somewhere, and on a PC that’s mostly system RAM.
| PC | gpt-oss-120b |
|---|---|
| 32 or 64 GB of RAM, any GPU | Too big for the app’s fit rule |
| 96 GB of RAM | Tight. Loads with little else running |
| 128 GB of RAM | Runs well |
| 128 GB of RAM plus a 24 to 32 GB card | Runs well, faster, with experts in RAM |
A bigger card speeds it up, because the parts that work on every word sit on the GPU. It doesn’t shrink the RAM you need for the rest.
What goes wrong, and the fix
| Symptom | Check | Fix |
|---|---|---|
| PC slows to a crawl on load | RAM in use by other apps | Close browsers and games; 16 GB is the floor |
| Now shows Vulkan or CPU on an NVIDIA card | Whether the NVIDIA download finished | Download it in Settings → GPU performance and restart |
| Out of memory in long chats | Context window in Model settings | Lower it, or start a new chat |
| Can’t find a 6 or 8 GB file | gpt-oss ships its experts in 4-bit | Take a file near 12 GB; there’s no small one |
| Image questions do nothing | gpt-oss is text only | Use Qwen 3.5 9B or Bonsai 2 27B from Models > Text |
For out-of-memory errors from CUDA itself, see how to fix CUDA out of memory on Windows.
Need a bigger model than your PC holds?
If you want a large open model without building a 128 GB workstation, Off Grid AI Boost is a flat-price API for open models on our GPUs. It names Qwen 3.8 and DeepSeek V4.1 Flash, and hasn’t listed gpt-oss, so don’t plan on the 120B there.
Boost starts at $4.99 a month for the first cohort, with regular prices from $9.99 to $199.99. We don’t store your prompts or train on them. That’s our promise today, and the full policy comes with launch. Usage limits aren’t published yet.
Boost starts in January; pre-order to lock in half price. API keys arrive by January 2027.
FAQ
What GPU do I need for gpt-oss-20b on Windows? None is required. A 16 GB card such as an RTX 5060 Ti 16 GB, 4060 Ti 16 GB, 5070 Ti or 5080 holds the whole 12 GB file. Smaller cards keep part of it in system RAM, which works because it’s a mixture-of-experts model.
How much RAM does gpt-oss-20b need on a Windows PC? 16 GB is the floor, and Off Grid AI Desktop marks it Tight there. 24 GB or more runs well.
Can a Windows PC run gpt-oss-120b? Only a big one. The file is 63.4 GB. Off Grid AI Desktop’s fit rule needs about 96 GB of system RAM to load it, and 128 GB to run it well. A large GPU makes it faster but doesn’t replace the RAM.
Does it work on AMD or Intel graphics? Yes. Off Grid AI Desktop uses Vulkan on AMD and Intel GPUs on Windows. NVIDIA cards use the optional CUDA download.
Is gpt-oss in the app’s model list? Not yet. Search for gpt-oss-20b in the Models screen and download it from the ggml-org or Unsloth upload on Hugging Face.
Related guides
- How to Run gpt-oss-20b on a 16GB Laptop in 2026
- How to Use Your NVIDIA GPU for Local AI on Windows in 2026
- How to Run LLMs Locally on Your Windows PC in 2026
- How to Run Qwen 3.8 27B on Windows in 2026
Check your RAM and VRAM in Task Manager under Performance, then download Off Grid AI Desktop and load gpt-oss-20b.