Model Overview

Model Architecture: gpt-oss-120b
- Input: Text
- Output: Text
Supported Hardware Microarchitecture: AMD MI350/MI355
ROCm: 7.0
Operating System(s): Linux
Inference Engine: vLLM
Model Optimizer: AMD-Quark
- Weight quantization: OCP MXFP4, Static
- Activation quantization: OCP MXFP4, Dynamic
Calibration Dataset: Pile

This model was built with gpt-oss-120b model by applying AMD-Quark for MXFP4 quantization.

Model Quantization

The model was quantized from openai/gpt-oss-120b using AMD-Quark. The weights are quantized MXFP4 and activations were quantized to FP8.

Quantization scripts:

cd Quark/examples/torch/language_modeling/llm_ptq/
exclude_layers="*lm_head *self_attn* *router*"

python3 internal_scripts/quantize_quark.py \
    --model_dir openai/gpt-oss-120b \
    --quant_scheme w_mxfp4_a_fp8 \
    --exclude_layers $exclude_layers \
    --num_calib_data 512 \
    --output_dir amd/gpt-oss120b-w-mxfp4-a-fp8 \
    --model_export hf_format \
    --multi_gpu

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend.

Evaluation

The model was evaluated on AIME25 and GPQA Diamond benchmarks with low reasoning effort.

Accuracy

Benchmark	gpt-oss-120b	gpt-oss120b-w-mxfp4-a-fp8(this model)	Recovery
AIME25	65.25	67.12	102.87%
GPQA	51.67	53.42	103.39%

Reproduction

The results of AIME25 and GPQA Diamond were obtained using gpt_oss.evals with low effort setting, and vLLM docker rocm/vllm-private:rocm7.0_ubuntu_22.04_vllm_0.10.1_instinct_gptoss_wmxfp4_afp8_20251030.

Launching server

vllm serve amd/gpt-oss120b-w-mxfp4-a-fp8 \
  --tensor_parallel_size 2 \
  --gpu-memory-utilization 0.90 \
  --no-enable-prefix-caching \
  --max-num-batched-tokens 1024

Evaluating model in a new terminal

python -m gpt_oss.evals --model /shareddata/amd/gpt-oss120b-w-mxfp4-a-fp8 --eval aime25,gpqa --reasoning-effort low --n-threads 128

jamiefutch/gpt-oss-120b-w-mxfp4-a-fp8