Can I run gpt-oss-20b on the GeForce RTX 5070?
Runs slowlyMXFP4 keeps attention on the GPU (2.0 GB) and its experts in 11.5 GB of system RAM: Runs slowly, est. 18.5 tok/s (12.9–24.0, theoretical estimate ±30%).
12 GB VRAM · 672.0 GB/s · FP16 30.9 TFLOPS. Assumes 32 GB of DDR5-5600 system RAM, Windows with this GPU driving the display, 8k context and an f16 KV cache. The GeForce RTX 5070 Ti Laptop has the same memory setup (12 GB, 672.0 GB/s), so its generation-speed estimate is the same; only prompt processing differs.
Every quant of gpt-oss-20b on the GeForce RTX 5070
| Quant | File size | Verdict | Speed | Memory | Runs as | Notes |
|---|---|---|---|---|---|---|
| MXFP4baseline | 12.11 GB | Runs slowly | est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 2.0 / 12.0 GB + 11.5 GB RAM | MoE experts in RAM | Experts run from system RAM — Runs well needs 30 tok/s or more on this path |
Reasoning models spend extra tokens thinking, so their speed thresholds are 1.5× stricter (30 / 12 / 3 tok/s).
Where the memory goes at MXFP4
- Weights
- 0.7 GB
- KV cache
- 0.2 GB
- Compute buffer
- 0.6 GB
- OS reserve
- 0.6 GB
- Free
- 10.0 GB
- Weights in system RAM
- 11.5 GB
Context length vs. KV cache
| Context | KV f16 | KV q8_0 | KV q4_0 |
|---|---|---|---|
| 4k | 12.7 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 12.7 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 12.7 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% |
| 8k | 12.9 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 12.8 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 12.7 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% |
| 16k | 13.2 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 13.0 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 12.9 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% |
| 32k | 13.7 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 13.3 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 13.1 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% |
| 64k | 14.8 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 14.1 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 13.7 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% |
| 128k | 17.0 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 15.5 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% | 14.7 GBRuns slowly est. 18.5 tok/s12.9–24.0theoretical estimate ±30% |
S = Runs great · A = Runs well · B = Runs slowly · F = Won't run
At MXFP4 the verdict stays Runs slowly up to 128k context with an f16 KV cache, and up to 128k with q8_0.
Estimated speed
- Generation
- est. 18.5 tok/s (12.9–24.0, theoretical estimate ±30%)
- Prompt processing
- Prompt-processing estimates only apply when the whole model runs on the GPU.
Why this verdict: Experts run from system RAM — Runs well needs 30 tok/s or more on this path
Measured on this exact combination
No public measurement for this exact combination yet; the numbers above are estimates.
If this is not enough
A smaller model that runs well on the GeForce RTX 5070
Runs greatQwen3 14B: Runs great, est. 45.5 tok/s (40.0–50.9, calibrated estimate ±12%)
Keep only the experts in system RAM
--n-cpu-moe 24 puts the experts of every layer in RAM (this estimate); 5 is the smallest value that still fits, and fewer layers on the CPU means faster.
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 --n-cpu-moe 24 -fa on- -ngl all loads every layer on the GPU.
- --n-cpu-moe 24 keeps the experts of all 24 layers in system RAM, which is what this estimate assumes.
- With 5 the model still fits; fewer layers on the CPU means faster generation.
- -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.
Ollama has no expert-offload flag, so this mode is llama.cpp only.
Related pages
Guide: How much VRAM do local LLMs need? 8–32 GB tiers (2026)
Other models on the GeForce RTX 5070
gpt-oss-20b on other hardware
Frequently asked questions
Can I run gpt-oss-20b on the GeForce RTX 5070?
MXFP4 keeps attention on the GPU (2.0 GB) and its experts in 11.5 GB of system RAM: Runs slowly, est. 18.5 tok/s (12.9–24.0, theoretical estimate ±30%). MXFP4 needs 12.9 GB at 8k context; this setup has 11.4 GB of usable memory and 28.0 GB of free system RAM.
How much context can gpt-oss-20b use on the GeForce RTX 5070?
At MXFP4 the verdict stays Runs slowly 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 slowly, est. 18.5 tok/s (12.9–24.0, 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.