← CanIRunLocal · All hardware · All models
Yes. gpt-oss 20B (MoE, MXFP4) fits in fast memory on DGX Spark 128GB (NVIDIA, ASUS, Dell, Lenovo, HP, MSI, Acer, Gigabyte) at MXFP4 (15.1 GB needed, 108.8 GB usable), about 65 tok/s (estimate).
Hardware: DGX Spark 128GB (NVIDIA, ASUS, Dell, Lenovo, HP, MSI, Acer, Gigabyte) · Model: gpt-oss 20B (MoE, MXFP4)
DGX Spark 128GB (NVIDIA, ASUS, Dell, Lenovo, HP, MSI, Acer, Gigabyte) has 128 GB of unified memory and 273 GB/s of memory bandwidth. DGX OS keeps about 15% of unified memory, so Ollama can use 108.8 GB. gpt-oss 20B (MoE, MXFP4) (20.9B, 3.6B active, OpenAI) at MXFP4 is a 14 GB download and needs 15.1 GB with headroom for about 8k tokens.
| Best quant | MXFP4 |
|---|---|
| Download size | 14 GB |
| Memory needed | 15.1 GB |
| Usable fast memory | 108.8 GB |
| Fit | fits in fast memory |
| Estimated speed | ~65 tok/s (estimate) |
| Command | ollama run gpt-oss:20b |
Estimates, not benchmarks. Model sizes are ollama.com download sizes; memory needed adds headroom for ~8k tokens of context; speeds are rough figures from memory bandwidth. Data updated 2026-10-23.
A project by OnDevice Labs · llms.txt