Can an RTX 3090 run gpt-oss-20b?
Yes: an RTX 3090 runs gpt-oss-20b at 207.2 tok/s at Q4_K_M, measured on llama.cpp with a 166-token prompt in October 2026.
1 signed run on llm-speed. Single-request decode speed in tokens per second; the short-chat workload unless noted. This run was measured by llm-speed (@llmspeed). Runs measured by llm-speed are credited to @llmspeed.
Measured tok/s by quant and runtime
| Quant | Runtime | Build | Decode tok/s (best run) | Workload | Runs | Source |
|---|---|---|---|---|---|---|
| Q4_K_M | llama.cpp | commit ea0c4c2 | 207.2 | chat-short | 1 | Measured by llm-speed@llmspeed |
Fastest quant that fits: Q4_K_M on llama.cpp, 207.2 tok/s (measured by llm-speed; it fits because it ran on this hardware). Quants without a run are not listed.
VRAM: not recorded for these runs. The VRAM calculator gives an estimate (predicted, not measured).
Every measured row
| Run | Date (UTC) | Workload | Quant | Runtime / build | Prompt tokens | Output tokens | Decode tok/s | Prefill tok/s | TTFT ms | VRAM GiB | Source |
|---|---|---|---|---|---|---|---|---|---|---|---|
| r_nakw1nq6clw | 2026-10-17 | chat-short | Q4_K_M | llama.cpp · commit ea0c4c2 | 166 | 256 | 207.2 | 658.3 | 252 | – | Measured by llm-speed@llmspeed |
| r_nakw1nq6clw | 2026-10-17 | chat-long | Q4_K_M | llama.cpp · commit ea0c4c2 | 3188 | 998 | 199.8 | 4606.3 | 692 | – | Measured by llm-speed@llmspeed |
| r_nakw1nq6clw | 2026-10-17 | agent-trace | Q4_K_M | llama.cpp · commit ea0c4c2 | 3287 | 840 | 202.2 | 9892.7 | 224 | – | Measured by llm-speed@llmspeed |
| r_nakw1nq6clw | 2026-10-17 | concurrent-decode | Q4_K_M | llama.cpp · commit ea0c4c2 | 1024 | 256 | 205.1 | – | – | – | Measured by llm-speed@llmspeed |
Questions
Can an RTX 3090 run gpt-oss-20b?
Yes: an RTX 3090 runs gpt-oss-20b at 207.2 tok/s at Q4_K_M, measured on llama.cpp with a 166-token prompt in October 2026.
How many tokens per second does gpt-oss-20b get on an RTX 3090?
Decode speed measured by llm-speed, by quantization and runtime: Q4_K_M on llama.cpp: 207.2 tok/s. Based on 1 run measured by llm-speed; single-request chat-short unless noted.
What is the fastest quant of gpt-oss-20b that fits on an RTX 3090?
Q4_K_M on llama.cpp was the fastest measured setting at 207.2 tok/s; it fit because it ran. Other quants are listed only where a run measured them.
How much VRAM does gpt-oss-20b use on an RTX 3090?
Memory use was not recorded for these runs. The VRAM calculator gives an estimate, not a measurement.