Hardware guide · gpt-oss
What hardware do you need to run gpt-oss 20B?
gpt-oss 20B needs about 13.2 GB to run at Q4_K_M, so the lightest hardware that runs it is Nvidia GeForce RTX 4060 Ti (16GB). 17 of 43 devices tested can run it.
- Needs (Q4_K_M)
- ~13.2 GB
- Devices that run it
- 17
- Too small
- 26
- Lightest that runs it
- ~15 GB
Runs at Q4_K_M using ~13.2 GB of ~15 GB usable. You have room for Q8_0 for higher quality.
What kind of hardware runs gpt-oss 20B
Of the 17 devices that run it, here is the split by hardware class.
Hardware that runs gpt-oss 20B
Ranked by usable memory, lightest first. Prices are approximate street prices for the device itself (a GPU is the card alone; a Mac is the whole machine). Tok/s is a bandwidth estimate; see methodology.
- YesNvidia GeForce RTX 4060 Ti (16GB) from $499~15 GB usable · uses ~13.2 GB
- YesNvidia GeForce RTX 4080 (16GB) from $1,199~15 GB usable · uses ~13.2 GB
- YesApple M4 (24GB) from $1,299~16 GB usable · uses ~13.2 GB
- YesApple M4 Pro (24GB) from $1,999~16 GB usable · uses ~13.2 GB
- YesApple M5 (32GB) from $1,299~21 GB usable · uses ~13.2 GB
- YesNvidia GeForce RTX 4090 (24GB) from $1,599~23 GB usable · uses ~13.2 GB
- YesNvidia GeForce RTX 3090 (24GB) from $1,499~23 GB usable · uses ~13.2 GB
- YesAMD Radeon RX 7900 XTX (24GB) from $999~23 GB usable · uses ~13.2 GB
- Yes32GB RAM Laptop (CPU/iGPU only) from $1,100~28 GB usable · uses ~13.2 GB
- YesNvidia GeForce RTX 5090 (32GB) from $1,999~31 GB usable · uses ~13.2 GB
- YesApple M4 Pro (48GB) from $2,399~32 GB usable · uses ~13.2 GB
- YesApple M5 Pro (48GB) from $2,199~32 GB usable · uses ~13.2 GB
- YesApple M4 Max (64GB) from $3,499~48 GB usable · uses ~13.2 GB
- YesApple M4 Max (128GB) from $3,499~96 GB usable · uses ~13.2 GB
- YesAMD Ryzen AI Halo (128GB) from $3,999~96 GB usable · uses ~13.2 GB
- YesApple M5 Max (128GB) from $3,599~96 GB usable · uses ~13.2 GB
- YesApple M3 Ultra (256GB) from $3,999~192 GB usable · uses ~13.2 GB
Too small for gpt-oss 20B
See the full gpt-oss 20B memory breakdown, or compare it against other models.
Sources
-
huggingface.co · 2 sources
-
ollama.com · 2 sources
Memory at Q4_K_M (13.2 GB = 11.28 GB weights + KV cache + overhead), Catalog updated 2026-10-05. See methodology.