Can the RTX 3080 Ti run gpt-oss-20b?
Not fully, but it's usable. gpt-oss-20b needs about 13.8 GB at Q4_K_M and the RTX 3080 Ti has 12 GB - yet as a mixture-of-experts model it reads only ~3.6B parameters per token, so you can keep attention on the GPU and the experts in system RAM (llama.cpp --n-cpu-moe) and still get interactive speeds with 32 GB+ of RAM.
Every quantization
gpt-oss-20b (21.5B, 3.6B active) on 12 GB
| Quant | Weights | Needs | Fit | Est. speed |
|---|---|---|---|---|
| Q4_K_M | 12 GB | 13.8 GB | Too big | - |
| Q5_K_M | 15 GB | 16.8 GB | Too big | - |
| Q8_0 | 23 GB | 24.8 GB | Too big | - |
| FP16 | 43 GB | 44.8 GB | Too big | - |
Estimates. Needs = weights + 1.8 GB for KV cache and runtime at modest context. Speed is single-stream decode scaled from the RTX 3080 Ti's 912 GB/s bandwidth (XiongjieDai/GPU-Benchmarks-on-LLM-Inference). Long contexts need more memory and run slower.
Run it with expert offload
Attention on the GPU, experts in system RAM
llama-server -hf unsloth/gpt-oss-20b-GGUF -ngl 99 --n-cpu-moe 99Lower --n-cpu-moe step by step until the RTX 3080 Ti's 12 GB is nearly full - each expert layer moved back to the GPU speeds things up. Needs about 8 GB+ of free system RAM.
What to do instead
Two ways forward
Cheapest cards that run it
- NVIDIA Tesla P40 - 24 GB, ~$159
- Intel Arc A770 16GB - 16 GB, ~$229
- AMD Radeon RX 7600 XT - 16 GB, ~$309
- NVIDIA Tesla T4 - 16 GB, ~$349
Models that fit the RTX 3080 Ti
- Qwen3 14B - 14B, 8 GB at Q4
- Gemma 4 12B - 12B, 7 GB at Q4
- Qwen3.5 9B - 9.7B, 6 GB at Q4
- Llama 3.1 8B - 8B, 5.5 GB at Q4
Compare
Same model, other hardware - and other models on this card
gpt-oss-20b on other hardware
- RTX 5090~341 tok/s
- RTX 5090 D~341 tok/s
- RTX 6000D~299 tok/s
- RTX 4090~208 tok/s
- Radeon RX 7900 XTX~183 tok/s
- RTX 5080~183 tok/s
- RTX 3090~182 tok/s
- RTX 5080 Mobile~171 tok/s
Also runs on the RTX 3080 Ti
- Qwen3 14B~65 tok/s
- Gemma 4 12B~75 tok/s
- Qwen3.5 9B~87 tok/s
- Llama 3.1 8B~95 tok/s
- Mistral-7B-Instruct-v0.2~131 tok/s
- Qwen3.5 4B~174 tok/s
- Qwen3-4B-Instruct-2507~261 tok/s
- NVIDIA Nemotron 3 Nano 4B~261 tok/s
Frequently asked
Quick answers
Can the NVIDIA RTX 3080 Ti run gpt-oss-20b?
Not fully, but it's usable. gpt-oss-20b needs about 13.8 GB at Q4_K_M and the RTX 3080 Ti has 12 GB - yet as a mixture-of-experts model it reads only ~3.6B parameters per token, so you can keep attention on the GPU and the experts in system RAM (llama.cpp --n-cpu-moe) and still get interactive speeds with 32 GB+ of RAM.
How much VRAM does gpt-oss-20b need?
About 12 GB of weights at Q4_K_M, 23 GB at Q8_0 and 43 GB at FP16. Add roughly 1.8 GB for the KV cache and runtime at modest context lengths.
How fast is gpt-oss-20b on the RTX 3080 Ti?
It doesn't fit in the RTX 3080 Ti's memory, so speed depends on how many layers you offload to system RAM - typically single-digit tokens per second or lower.