Can I run gpt-oss-120b on the GeForce RTX 5090?
Won't runMXFP4 needs 63.6 GB, 32.2 GB more than the 31.4 GB of usable memory. With 64 GB of system RAM: Runs well, est. 35.5 tok/s (24.8–46.1, theoretical estimate ±30%).
32 GB VRAM · 1,792.0 GB/s · FP16 104.8 TFLOPS. Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache.
Verdict by system RAM
This page assumes 32 GB of system RAM. For this pair, the verdict depends on how much system RAM you have.
- 16 GBWon't runMXFP4 doesn't fit
- 32 GBWon't runMXFP4 doesn't fit(used on this page)
- 64 GBRuns wellest. 35.5 tok/s (24.8–46.1, theoretical estimate ±30%)(used for the commands below)
- 96 GBRuns wellest. 35.5 tok/s (24.8–46.1, theoretical estimate ±30%)
- 128 GBRuns wellest. 35.5 tok/s (24.8–46.1, theoretical estimate ±30%)
Every quant of gpt-oss-120b on the GeForce RTX 5090
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| MXFP4baseline | 63.39 GB | Won't run | — | Needs about 63.6 GB; 31.4 GB of VRAM and 28.0 GB of RAM are free | — | — |
Models with a thinking mode spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s). With thinking turned off, a model may do one grade better.
Where the memory goes at MXFP4
The memory, context and speed figures below assume 64 GB of system RAM (the configuration from the RAM hint above); with 32 GB nothing fits.
- Weights
- 27.1 GB
- KV cache
- 0.3 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Free
- 3.4 GB
- Weights in system RAM
- 36.3 GB
System RAM, context length and KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 63.4 GBRuns well est. 37.1 tok/s26.0–48.3theoretical estimate ±30% | 63.4 GBRuns well est. 37.2 tok/s26.1–48.4theoretical estimate ±30% | 63.3 GBRuns well est. 37.3 tok/s26.1–48.4theoretical estimate ±30% |
| 8k | 63.6 GBRuns well est. 35.5 tok/s24.8–46.1theoretical estimate ±30% | 63.5 GBRuns well est. 37.1 tok/s26.0–48.3theoretical estimate ±30% | 63.4 GBRuns well est. 37.2 tok/s26.1–48.4theoretical estimate ±30% |
| 16k | 64.0 GBRuns well est. 35.2 tok/s24.6–45.7theoretical estimate ±30% | 63.7 GBRuns well est. 35.5 tok/s24.8–46.1theoretical estimate ±30% | 63.6 GBRuns well est. 37.1 tok/s26.0–48.3theoretical estimate ±30% |
| 32k | 64.8 GBRuns well est. 34.6 tok/s24.2–45.0theoretical estimate ±30% | 64.2 GBRuns well est. 35.1 tok/s24.6–45.7theoretical estimate ±30% | 63.9 GBRuns well est. 35.4 tok/s24.8–46.1theoretical estimate ±30% |
| 64k | 66.3 GBRuns well est. 32.3 tok/s22.6–41.9theoretical estimate ±30% | 65.2 GBRuns well est. 34.5 tok/s24.2–44.9theoretical estimate ±30% | 64.5 GBRuns well est. 35.1 tok/s24.6–45.6theoretical estimate ±30% |
| 128k | 69.3 GBRuns slowly est. 28.4 tok/s19.9–36.9theoretical estimate ±30% | 67.0 GBRuns well est. 31.0 tok/s21.7–40.3theoretical estimate ±30% | 65.8 GBRuns well est. 33.1 tok/s23.2–43.1theoretical estimate ±30% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
With 64 GB of system RAM: At MXFP4 the verdict stays Runs well up to 64k context with an f16 KV cache, and up to 128k with q8_0.
Estimated speed
- Generation
- With 64 GB of system RAM: est. 35.5 tok/s (24.8–46.1, theoretical estimate ±30%)
- Prompt processing
- Prompt-processing estimates only apply when the whole model runs on the GPU.
Why this verdict: Not fully on the GPU — Runs great needs the whole model in GPU memory
Measured on this exact combination
Public benchmarks we calibrate against. Their conditions (context, backend, flags) can differ from the estimates above.
| Hardware | Quant | Backend | Context | Prompt (tok/s) | Generation (tok/s) | Flags | Source | Measured |
|---|---|---|---|---|---|---|---|---|
| GeForce RTX 5090 | MXFP4 | llama.cpp | 4k | — | 9.6 | --n-cpu-moe 21, DDR4 server | hardware-corner.net | 2026-08-09 |
If this is not enough
A smaller model that runs well on the GeForce RTX 5090
Runs wellKolibri 1: Runs well, est. 83.3 tok/s (58.3–108.3, theoretical estimate ±30%)
Same GPU with 64 GB of system RAM
Runs wellMXFP4: Runs well, est. 35.5 tok/s (24.8–46.1, theoretical estimate ±30%)
How to run it
Commands for the MXFP4 file. Both tools download from Hugging Face on first run.
These commands assume 64 GB of system RAM (the configuration from the RAM hint above). For a different amount of RAM, pick it under “System RAM, context length and KV cache” above to see the commands for it.
File: gpt-oss-120b-MXFP4.gguf (63.39 GB) from ggml-org/gpt-oss-120b-GGUF.
llama-server -hf ggml-org/gpt-oss-120b-GGUF:MXFP4 -c 8192 -ngl all --n-cpu-moe 21 -fa on- -ngl all loads every layer on the GPU.
- --n-cpu-moe 21 keeps the experts in the first 21 layers in system RAM; the remaining 15 layers with experts keep them on the GPU. This estimate assumes the same split.
- If loading fails with an out-of-memory error, raise the value to 22, or to 23 if that isn't enough.
- llama.cpp b7410 (December 2025) and later split the model in a similar way on their own if you leave out -ngl and --n-cpu-moe.
- -fa on enables flash attention (needed for KV cache quantization).
- --cache-type-k q8_0 --cache-type-v q8_0 shrinks the KV cache to 0.16 GB at 8k context.
PowerShell (quit the Ollama tray app first):
$env:OLLAMA_CONTEXT_LENGTH="8192"; ollama serve
ollama run hf.co/ggml-org/gpt-oss-120b-GGUF:MXFP4- Ollama defaults to 4k context on GPUs under 24 GB. Set OLLAMA_CONTEXT_LENGTH when starting the server, or /set parameter num_ctx inside the chat.
- Ollama 0.30 and later automatically keep part of the experts in system RAM to fit your VRAM. Older versions split whole layers instead, which is much slower.
/set parameter num_ctx 8192
Related pages
Guide: How much VRAM do local LLMs need? GPU requirements for 8–32 GB (2026)
Other models on the GeForce RTX 5090
gpt-oss-120b on other hardware
Frequently asked questions
Can I run gpt-oss-120b on the GeForce RTX 5090?
MXFP4 needs 63.6 GB, 32.2 GB more than the 31.4 GB of usable memory. With 64 GB of system RAM: Runs well, est. 35.5 tok/s (24.8–46.1, theoretical estimate ±30%). MXFP4 needs 63.6 GB at 8k context; this setup has 31.4 GB of usable memory and 28.0 GB of free system RAM.
How much context can gpt-oss-120b use on the GeForce RTX 5090?
With 64 GB of system RAM: At MXFP4 the verdict stays Runs well up to 64k context with an f16 KV cache (2.4 GB of KV) and up to 128k with a q8_0 KV cache (2.6 GB). The model supports up to 128k.
Which quant should I use, and how fast is it?
No tracked quant runs with 32 GB of RAM; with 64 GB, MXFP4 reaches Runs well, est. 35.5 tok/s (24.8–46.1, theoretical estimate ±30%).
Speeds are estimates from memory bandwidth, calibrated against public benchmarks, and each one comes with an error band and a confidence label. Real results vary with drivers, backend, context length and thermals.