Hardware to run gpt-oss-120b
The model that sells 128 GB machines: about 63 GB in any quantisation and only 5.1B active parameters.
- Developer
- OpenAI
- Parameters
- 117B (active 5.1B)
- Max context
- 128K tokens
- Licence
- Apache 2.0
Model card on Hugging FaceGGUF filesOllama: ollama run gpt-oss:120b
Memory needed
Factual dataWeights plus context cache plus about 1.5 GB of runtime buffers, by quantisation and context length.
| Quantisation | 8K tokens | 32K tokens | 128K tokens |
|---|---|---|---|
| Q4_K_M | 65 GB | 65 GB | 69 GB |
| Q8_0 | 65 GB | 66 GB | 70 GB |
| FP16 | 67 GB | 68 GB | 72 GB |
Hardware that runs it
Estimate8 platforms can run gpt-oss-120b at Q4_K_M with a 32K-token context. Needs about 65 GB.
Cheapest that runs it
Fastest that runs it
Model weightsContext cache (KV)RuntimeUsable limit (colour = verdict)
RunsRuns comfortably (7)
| Platform | Smallest configuration | Memory | Speed ceiling | Price from |
|---|---|---|---|---|
| RTX PRO 6000 | 96 GB | ≤ 654 tok/s | $8,565 | |
| Mac Studio M5 Max | 128 GB | ≤ 224 tok/s | — | |
| Ryzen AI Max+ PRO 495 | 192 GB | ≤ 100 tok/s | $7,399 | |
| Jetson AGX Thor | 128 GB | ≤ 100 tok/s | $5,499 | |
| DGX Spark / GB10 | 128 GB | ≤ 100 tok/s | $5,000 | |
| Ryzen AI Max+ 395 | 128 GB | ≤ 93 tok/s | $2,459 | |
| Ryzen AI 9 HX 370 | 96 GB | ≤ 33 tok/s | — |
Runs, tightRuns, but tight (1)
| Platform | Smallest configuration | Memory | Speed ceiling | Price from |
|---|---|---|---|---|
| Mac Studio M5 Ultra | 96 GB | ≤ 438 tok/s | $5,499 |
Partly in system RAMOnly with part of the model in system RAM (19)
| Platform | Smallest configuration | Memory | Speed ceiling | Price from |
|---|---|---|---|---|
| RTX PRO 5000 | 48 GB | — | $4,500 | |
| Radeon AI PRO R9700 | 32 GB | — | $1,550 | |
| Arc Pro B70 | 32 GB | — | $1,268 | |
| RTX 5090 | 32 GB | — | $4,329 | |
| RTX PRO 4500 | 32 GB | — | $2,500 | |
| RX 7900 XTX | 24 GB | — | $1,358 | |
| Arc Pro B60 | 24 GB | — | $599 | |
| RTX 3090 | 24 GB | — | $1,340 | |
| RTX 4090 | 24 GB | — | $1,599 | |
| RTX PRO 4000 | 24 GB | — | $1,500 | |
| RX 9060 XT 16 GB | 16 GB | — | $349 | |
| RX 9070 XT | 16 GB | — | $759 | |
| Arc Pro B50 | 16 GB | — | $349 | |
| RTX 5060 Ti 16 GB | 16 GB | — | $789 | |
| RTX 5070 Ti | 16 GB | — | $1,190 | |
| RTX 5080 | 16 GB | — | $1,830 | |
| Arc B580 | 12 GB | — | $249 | |
| RTX 3060 12 GB | 12 GB | — | $329 | |
| RTX 5070 | 12 GB | — | $1,010 |
Not enough memoryNot enough memory (4)
| Platform | Smallest configuration | Memory | Speed ceiling | Price from |
|---|---|---|---|---|
| Mac mini M5 Pro | 64 GB | — | $1,699 | |
| Jetson AGX Orin 64GB | 64 GB | — | $1,999 | |
| Mac mini M6 | 32 GB | — | $899 | |
| Jetson Orin Nano Super | 8 GB | — | $399 |
Not supportedNot supported (2)
| Platform | Smallest configuration | Memory | Speed ceiling | Price from |
|---|---|---|---|---|
| RK3588 (Orange Pi 5 Plus and others) | 32 GB | — | — | — |
| Raspberry Pi AI HAT+ 2 | 8 GB | — | — | $200 |
Theoretical maximum for generating text: memory bandwidth divided by the bytes read per token. Real runs usually reach about half to three quarters of it.