GeForce RTX 4060 · gpt-oss-20b · MXFP4
Only with CPU offload. gpt-oss-20b at MXFP4 does not fit in the 8 GB on GeForce RTX 4060, and a 32 GB system RAM pool runs it slowly, roughly 13–19 tokens per second.
| Piece | Value | Note |
|---|---|---|
| Weights | 11.3 GiB | gpt-oss-20b-MXFP4.gguf |
| KV cache | 0.19 GiB | 8,192 tokens |
| Engine overhead | 0.50 GiB | Flat 0.5 GiB allowance |
| 5% headroom | 0.60 GiB | On the sum above |
| Total | 12.6 GiB | Offload |
| Fast memory | 8.00 GiB | 8 GB VRAM |
| Speed | 13–19 tok/s | Active-expert bandwidth range |
ollama run gpt-oss:20b
GeForce RTX 4060 · gpt-oss-20b ·GGUF file · Methodology
FAQ
How was this fit calculated?
Weights 11.3 GiB, KV cache 0.19 GiB at 8,192 tokens, engine overhead 0.50 GiB, then 5% headroom, for 12.6 GiB. Fast memory available is 8.00 GiB.
What if I change the quant or the context?
This page is MXFP4 at 8,192 tokens. A smaller quant or a shorter context lowers the total. The homepage calculator changes both without a new page.
Can I trust the tokens per second?
The range 13–19 tok/s is a bandwidth roofline, 55% to 80% efficient. This MoE model only reads the active experts, using an architecture-derived fraction of the file. Measure your own box before you buy around it.
What do I type?
ollama run gpt-oss:20b
Is GeForce RTX 4060 assumed to have extra RAM?
Yes. Discrete GPU pages assume 32 GB of system RAM. Offload uses whatever remains after 8 GB for Windows, and it is limited by the slower of DDR5-5600 and the PCIe link.