Can I run gpt-oss-20b on the GeForce RTX 4090?
Runs greatMXFP4 fits in VRAM with 11.1 GB to spare: Runs great, 225.2 tok/s measured at 2k context (1 run; 8k estimate 214.1 tok/s, 171.3–256.9, calibrated estimate ±20%).
24 GB VRAM · 1,008.0 GB/s · FP16 82.6 TFLOPS. Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache. The verdict and quant file above stay the same anywhere from 16 to 128 GB of system RAM. The GeForce RTX 3090 Ti has the same memory setup (24 GB, 1,008.0 GB/s), so its generation-speed estimate is the same; only prompt processing differs. Prompt processing: about 12,544 tok/s (7,526–17,561, rough estimate ±40%) here vs about 6,074 tok/s (3,645–8,504, rough estimate ±40%) on the GeForce RTX 3090 Ti.
Every quant of gpt-oss-20b on the GeForce RTX 4090
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| MXFP4baseline | 12.11 GB | Runs great | 225.2 tok/s8k estimate 214.1 tok/s (171.3–256.9, calibrated estimate ±20%)measured (1 run, 2k context) | 12.9 / 24.0 GB | Full GPU | — |
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
- Weights
- 11.5 GB
- KV cache
- 0.2 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Free
- 11.1 GB
0.63 GB of the 12.11 GB file stays off the graphics card: llama.cpp keeps the input embedding table in system RAM.
Context length vs. KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 12.1 GBRuns great est. 220.9 tok/s176.7–265.0calibrated estimate ±20% | 12.1 GBRuns great est. 224.1 tok/s179.3–269.0calibrated estimate ±20% | 12.0 GBRuns great est. 226.0 tok/s180.8–271.1calibrated estimate ±20% |
| 8k | 12.3 GBRuns great 225.2 tok/s8k estimate 214.1 tok/s (171.3–256.9, calibrated estimate ±20%)measured (1 run, 2k context) | 12.2 GBRuns great est. 220.4 tok/s176.3–264.5calibrated estimate ±20% | 12.1 GBRuns great est. 223.9 tok/s179.1–268.7calibrated estimate ±20% |
| 16k | 12.5 GBRuns great est. 201.8 tok/s161.4–242.1calibrated estimate ±20% | 12.3 GBRuns great est. 213.2 tok/s170.6–255.8calibrated estimate ±20% | 12.2 GBRuns great est. 219.9 tok/s175.9–263.9calibrated estimate ±20% |
| 32k | 13.1 GBRuns great est. 181.0 tok/s144.8–217.1calibrated estimate ±20% | 12.7 GBRuns great est. 200.2 tok/s160.1–240.2calibrated estimate ±20% | 12.5 GBRuns great est. 212.3 tok/s169.8–254.8calibrated estimate ±20% |
| 64k | 14.2 GBRuns great est. 150.0 tok/s120.0–180.0calibrated estimate ±20% | 13.4 GBRuns great est. 178.4 tok/s142.7–214.0calibrated estimate ±20% | 13.0 GBRuns great est. 198.6 tok/s158.9–238.3calibrated estimate ±20% |
| 128k | 16.4 GBRuns great est. 111.7 tok/s89.4–134.1calibrated estimate ±20% | 14.9 GBRuns great est. 146.5 tok/s117.2–175.8calibrated estimate ±20% | 14.1 GBRuns great est. 175.9 tok/s140.7–211.0calibrated estimate ±20% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At MXFP4 the verdict stays Runs great up to 128k context with an f16 KV cache, and up to 128k with q8_0.
Estimated speed
- Generation
- 225.2 tok/s measured at 2k context (1 run; 8k estimate 214.1 tok/s, 171.3–256.9, calibrated estimate ±20%)
- Prompt processing
- about 12,544 tok/s (7,526–17,561, rough estimate ±40%)
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 4090 | MXFP4 | llama.cpp | 2k | 8,078.0 | 225.2 | — | github.com | 2025-08-20 |
If this is not enough
A larger model that also runs well on the GeForce RTX 4090
Runs wellK2-Horizon MoVA 36B-A4B: Runs well, est. 82.6 tok/s (57.8–107.4, theoretical estimate ±30%)
How to run it
Commands for the MXFP4 file. Both tools download from Hugging Face on first run.
File: gpt-oss-20b-MXFP4.gguf (12.11 GB) from ggml-org/gpt-oss-20b-GGUF.
llama-server -hf ggml-org/gpt-oss-20b-GGUF:MXFP4 -c 8192 -ngl all -fa on- -ngl all loads every layer on the GPU.
- -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 0.11 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-20b-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.
/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 4090
gpt-oss-20b on other hardware
Frequently asked questions
Can I run gpt-oss-20b on the GeForce RTX 4090?
MXFP4 fits in VRAM with 11.1 GB to spare: Runs great, 225.2 tok/s measured at 2k context (1 run; 8k estimate 214.1 tok/s, 171.3–256.9, calibrated estimate ±20%). MXFP4 needs 12.3 GB at 8k context; this setup has 23.4 GB of usable memory and 28.0 GB of free system RAM.
How much context can gpt-oss-20b use on the GeForce RTX 4090?
At MXFP4 the verdict stays Runs great up to 128k context with an f16 KV cache (3.2 GB of KV) and up to 128k with a q8_0 KV cache (1.7 GB). The model supports up to 128k.
Which quant should I use, and how fast is it?
MXFP4 (12.11 GB) is the only tracked file: Runs great, 225.2 tok/s measured at 2k context (1 run; 8k estimate 214.1 tok/s, 171.3–256.9, calibrated estimate ±20%).
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.