GPT-OSS 120B
Hard reasoning and agent tasks on a high-end workstation. Plan on a 128 GB machine. Your laptop just flinched.
Model / OpenAI
Trust score 84 of 100. Yes, OSI-approved license.
Short answer (as of Oct 14, 2026): GPT-OSS 20B is a local AI model from OpenAI. Truly open? Yes, OSI-approved license. Business use: yes. OpenAI says it runs in 16 GB of memory; our estimate is about 17 GB, so 16 GB machines will be tight. Trust score 84 of 100.
As of Oct 14, 2026: license Apache License 2.0; business use yes; sign-up not needed; OpenAI says 16 GB of memory, our estimate about 17 GB; trust score 84/100. Sources.
Reasoning, tool use, and general chat.
Who makes it: OpenAI. The files come from the maker's own Hugging Face account.
License: Apache License 2.0. A standard open source license. You can use, change, and share it, even to make money. Keep the license notice.
Business use: Yes. Business use is allowed. Keep the license and notice files when you share it.
Training data: We have not found a full public data release. Not yet checked in depth.
Pick a computer with at least 32 GB of memory. On a Mac, that is the unified memory. On a PC, a graphics card with that much VRAM is fastest. A PC with only normal RAM will also work, just slower. This model uses a mix of experts. Only part of it works on each word, so it runs faster than its size suggests.
OpenAI says it runs in 16 GB of memory (source). Our estimate is about 17 GB. So 16 GB machines will be a tight fit. Close the 40 browser tabs first.
lmstudio-community/gpt-oss-20b-GGUF.Q4_K_M. It is the best mix of size and quality for most people.The files are GGUF builds that LM Studio can load. We checked that this repo exists on Oct 14, 2026. Full LM Studio guide.
Does it fit? Yes. Its memory window is big enough for Hermes.
hermes model. Choose LM Studio and press Enter to keep http://localhost:1234/v1.hermes to start.Yes, OSI-approved license. A standard open source license. You can use, change, and share it, even to make money. Keep the license notice.
Yes. Business use is allowed. Keep the license and notice files when you share it.
No. Download without an account.
OpenAI says it runs within 16 GB of memory. Our estimate is about 17 GB of RAM for a 4-bit file of about 12.5 GB. This is a rough guide, not a test. So a 16 GB machine will be a tight fit.
| Fact | What we found | Status |
|---|---|---|
| Maker | OpenAI | Verified Oct 14, 2026 |
| License | Apache License 2.0 | Verified Oct 14, 2026 |
| License plain | A standard open source license. You can use, change, and share it, even to make money. Keep the license notice. | Verified Oct 14, 2026 |
| Truly open | Yes, OSI-approved license | Verified Oct 14, 2026 |
| Business use | Yes. Business use is allowed. Keep the license and notice files when you share it. | Verified Oct 14, 2026 |
| Signup needed | No. Download without an account. | Verified Oct 14, 2026 |
| Parameters | 20.9 billion | Verified Oct 14, 2026 |
| Open training data | Not yet checked | We have not found a full public training data release. Not yet checked in depth. |
| Context window | 128K tokens | Verified Oct 14, 2026 |
| First posted | 2025-08-04 | Verified Oct 14, 2026 |
| Gguf files | lmstudio-community/gpt-oss-20b-GGUF | Verified Oct 14, 2026 |
| Hermes ok | Yes | Verified Oct 14, 2026 |
| Official memory | Runs within 16 GB of memory (OpenAI, MXFP4 build) | Verified Oct 14, 2026 |
If the maker's page and this page disagree, trust the maker. Tell us what we got wrong.
Newsletter
Which new models are worth your time, which ones are just hype, and the license gotchas before they bite. All the useful bits, none of the nerd-party small talk.
This opens our sign-up page on Beehiiv, our email provider. Confirm your email there to join.