Benchmark Report: gpt-oss-120b (mxfp4)
Date: 2026-06-03
Author: WITCHEER
Platform: NVIDIA GeForce RTX 5090 Benchmark Rig (capsule)
Model
| Field |
Value |
| Model |
gpt-oss-120b |
| Parameters |
116.83 B (dense) |
| Quantization |
mxfp4 |
| File size |
0.01 GiB |
| Engine |
llama.cpp (CUDA 12.8 (patched)) |
Hardware
| Component |
Spec |
| GPU |
NVIDIA GeForce RTX 5090 |
| CPU |
AMD Ryzen 5 9600 |
| RAM |
64GB DDR5-5600 |
| OS |
Ubuntu 26.04 LTS |
| CUDA |
12.8 (patched) |
Quality Benchmarks
All benchmarks use generative evaluation via llama-server chat completions. Multiple-choice tasks (MMLU, ARC, HellaSwag) use letter extraction instead of loglikelihood scoring -- results are internally consistent for model comparison but absolute scores may differ from logprob-based evaluations by 5-15%.
Summary
| Benchmark |
Score |
Metric |
| MMLU |
89.47% |
accuracy |
| ARC-Challenge |
95.00% |
accuracy |
| HellaSwag |
80.00% |
accuracy |
| HumanEval |
98.00% |
pass@1 |
| GSM8K |
97.00% |
exact_match |
MMLU Breakdown by Category
| Category |
Score |
Correct / Total |
| Stem |
97.22% |
35 / 36 |
| Humanities |
83.33% |
20 / 24 |
| Social Sciences |
88.46% |
23 / 26 |
| Other |
85.71% |
24 / 28 |
Speed Benchmarks
Measured with llama-bench. MoE experts offloaded to CPU (--n-cpu-moe 20), attention/active path on GPU (-ngl 99). Model is 59GB and does not fit 32GB VRAM; tuned to ~30GB VRAM with the rest in system RAM.
Prompt Processing (tokens/s)
| Context Length |
Speed |
+/-sigma |
| 128 |
124.5 |
33.1 |
| 512 |
484.4 |
37.3 |
| 2048 |
588.2 |
15.0 |
Generation (tokens/s)
| Metric |
Speed |
+/-sigma |
| tg128 |
46.5 |
0.4 |
Methodology
Evaluation Framework
Custom generative evaluators built for this rig. All benchmarks run through llama-server's /v1/chat/completions endpoint.
- Scoring: Generative evaluation (not loglikelihood)
- Thinking: enabled
- MCQ scoring: First valid letter extracted from response (A/B/C/D)
- Temperature: 0 (deterministic)
- Max tokens: 2,048 (MCQ/GSM8K); HumanEval 4,096, no stop sequences (reasoning models emit code after long inline reasoning)
- GPU offload: MoE experts of first 20 layers on CPU (
--n-cpu-moe 20); attention + active path on GPU. ~30GB of 32GB VRAM, remainder in RAM.
Benchmarked by WITCHEER on the RTX 5090 Benchmark Rig. Source: github.com/notwitcheer/llm-bench-rig/blob/main/reports/gpt-oss-120b-mxfp4.md. Dataset: huggingface.co/datasets/witcheer/rtx-5090-benchmarks/blob/main/reports/gpt-oss-120b-mxfp4.md.
