Can RTX 5090 32GB run gpt-oss 20B?
Measured results for gpt-oss 20B (20.91B MoE, 3.6B active) on RTX 5090 32GB, plus what fits and the commands to run it.
Short answer
Yes. gpt-oss 20B Q8_0 fits fully on RTX 5090 32GB at 16k context, using 12.5 of 31.0 GB usable, and generates about 374.7 tokens/s in our own run. As the context fills it slows down: 331 tok/s with 16k tokens already in context (88% of the empty-context speed), measured. For Q8_0, our bandwidth-only estimate was ~473 tok/s against a measured median of ~375 (26% too optimistic).
Measured generation speed by quantization
| Quant | Generation tok/s | Range | Prompt tok/s | Runs | Origin |
|---|---|---|---|---|---|
| Q8 class (median) | 374.7 | 374.7–374.7 | 18400 | 1 | includes our runs |
| Q4 class (median) | 298.5 | 282.3–393.3 | 12966 | 6 | includes our runs |
| Q4_K_M (our run, depth 0) | 393.3 | 3 repeats | 17630 | 3 | own measurement |
| Q8_0 (our run, depth 0) | 374.7 | 3 repeats | 18400 | 3 | own measurement |
Generation = tg128 (128 generated tokens, short context). Public rows are medians of community llama.cpp results (sources); our own runs use llama-bench b11658 on a rented GPU.
Speed vs context depth measured
Our own llama-bench runs on GeForce RTX 5090 (flash attention on, 3 repeats): generation and prompt speed with the given number of tokens already in the context.
| Quant | Context depth | Generation tok/s | vs depth 0 | Prompt tok/s |
|---|---|---|---|---|
| Q4_K_M | 0 | 393.3 | 100% | 17630 |
| Q4_K_M | 4k | 355 | 90% | 16548 |
| Q4_K_M | 16k | 345.2 | 88% | 13690 |
| Q8_0 | 0 | 374.7 | 100% | 18400 |
| Q8_0 | 4k | 339.3 | 91% | 17289 |
| Q8_0 | 16k | 331 | 88% | 14070 |
Alternatives
Less memory that still fits gpt-oss 20B
Smaller memory usually means a cheaper card or Mac (we do not list prices).
- RTX 5080 16GB: Q8_0, 15 GB usable, ~239 tok/s measured
- RTX 5070 Ti 16GB: Q8_0, 15 GB usable, ~224 tok/s measured
- RTX 5060 Ti 16GB: Q8_0, 15 GB usable, ~125 tok/s measured
Other models on RTX 5090 32GB, similar size
- LFM2.5 24B-A2B (hardware page): Q8_0, fits fully, ~343 tok/s (scaled from measured)
- Mistral Small 3.2 24B: Q8_0, fits fully, ~64 tok/s (measured)
- Gemma 4 26B-A4B: Q8_0, fits fully, ~217 tok/s (measured)
- Qwen3.6 27B: Q8_0, fits fully, ~52 tok/s (measured)
Memory and context limits
Q8_0 weights are 12.1 GB; the 16k f16 KV cache adds 0.41 GB. Largest context that fits in 31.0 GB usable:
| KV cache type | Max context |
|---|---|
| f16 | 128k |
| q8_0 | 128k |
| q4_0 | 128k |
Native limit 128k. Full KV cache table for gpt-oss 20B.
Copy-paste commands
The advisor's pick for this hardware: Q8_0, 16k context.
llama-server -hf unsloth/gpt-oss-20b-GGUF:Q8_0 -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/unsloth/gpt-oss-20b-GGUF:Q8_0
FAQ
How fast is gpt-oss 20B on RTX 5090 32GB?
About 374.7 tokens/s for generation (Q8_0 on llama.cpp CUDA, 128-token generation), 18400 tokens/s for prompt processing. Speed drops as the context fills; see the depth table.
How much memory does gpt-oss 20B need on RTX 5090 32GB?
The advisor's pick, Q8_0, is 12.1 GB of weights plus 0.41 GB of KV cache at 16k context (12.5 GB), against 31.0 GB usable.
Is the gpt-oss 20B speed on RTX 5090 32GB measured or estimated?
Measured: our own llama-bench run on a rented GeForce RTX 5090, including runs with 4k to 16k tokens already in context.
What can I use instead of RTX 5090 32GB for gpt-oss 20B?
With less memory, RTX 5080 16GB (~239 tok/s), RTX 5070 Ti 16GB (~224 tok/s), RTX 5060 Ti 16GB (~125 tok/s) also fit it fully.
What is the maximum context for gpt-oss 20B on RTX 5090 32GB?
With an f16 KV cache about 128k tokens, 128k with q8_0 and 128k with q4_0, bounded by the model's own limit.
Related
All models on RTX 5090 32GB · gpt-oss 20B on all hardware · VRAM tier · all measured pages · leaderboard
Compare: RTX 4090 24GB vs RTX 5090 32GB · RTX 3090 24GB vs RTX 5090 32GB · 2x RTX 3090 (48GB total) vs RTX 5090 32GB