Instructions to use amd/GLM-5.3-Quark-MXFP4-AttnFP8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use amd/GLM-5.3-Quark-MXFP4-AttnFP8 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="amd/GLM-5.3-Quark-MXFP4-AttnFP8")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("amd/GLM-5.3-Quark-MXFP4-AttnFP8")
model = AutoModelForCausalLM.from_pretrained("amd/GLM-5.3-Quark-MXFP4-AttnFP8", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use amd/GLM-5.3-Quark-MXFP4-AttnFP8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "amd/GLM-5.3-Quark-MXFP4-AttnFP8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "amd/GLM-5.3-Quark-MXFP4-AttnFP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/amd/GLM-5.3-Quark-MXFP4-AttnFP8

SGLang

How to use amd/GLM-5.3-Quark-MXFP4-AttnFP8 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "amd/GLM-5.3-Quark-MXFP4-AttnFP8" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "amd/GLM-5.3-Quark-MXFP4-AttnFP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "amd/GLM-5.3-Quark-MXFP4-AttnFP8" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "amd/GLM-5.3-Quark-MXFP4-AttnFP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use amd/GLM-5.3-Quark-MXFP4-AttnFP8 with Docker Model Runner:
```
docker model run hf.co/amd/GLM-5.3-Quark-MXFP4-AttnFP8
```

Model Overview

Model Architecture: GlmMoeDsaForCausalLM
- Input: Text
- Output: Text
Supported Hardware Microarchitecture: AMD MI350/MI355
ROCm: 7.1.1
PyTorch: 2.10.0+rocm7.1.1.lw.gitd9556b05
Transformers: 5.15.1
Operating System(s): Linux
Inference Engine: vLLM
Model Optimizer: AMD-Quark (V0.12)
- Quantized layers: self_attn, router experts, shared_experts
- self_attn: FP8 per-block
- Router experts and shared_experts: MXFP4

This model was built with GLM-5.3 model by applying AMD-Quark for MXFP4 quantization.

Model Quantization

The model was quantized from zai-org/GLM-5.3 using AMD-Quark. The weights and activations are quantized to MXFP4.

Quantization scripts:

cd Quark/examples/torch/language_modeling/llm_ptq/
python quantize_quark.py \
  --model_dir zai-org/GLM-5.3 \
  --output_dir amd/GLM-5.3-Quark-MXFP4 \
  --quant_scheme mxfp4 \
  --exclude_layers \
    "*self_attn*" \
    "*lm_head" \
    "*embed_tokens" \
    "*eh_proj" \
    "*mlp.gate" \
  --file2file_quantization \
  --keep_excluded_layers_as_original_model_state

Evaluation

The model was evaluated using vLLM, EvalScope, and lm-evaluation-harness.

Accuracy

Benchmark	zai-org/GLM-5.3	amd/GLM-5.3-Quark-MXFP4-AttnFP8 (this model)	Recovery
GSM8K (flexible-extract)	91.96	91.43	99.42%
GPQA Diamond (mean@8)	92.36	91.54	99.11%
AIME26 (mean@16)	98.12	98.96	100.86%
MMLU Pro (mean@1)	88.19	87.70	99.44%
SciCode, subproblem (mean@8)	56.14	56.60	100.82%
AA-LCR (mean@16)	72.44	71.69	98.96%
IFBench (mean@5)	67.33	66.67	99.02%

Baseline is zai-org/GLM-5.3 (native FP8). MMLU Pro is used in place of MMMU Pro as this model has no vision encoder.

Note: Results may vary from run to run.

Reproduction

The following results were obtained using vLLM, EvalScope 1.14.1, and lm-evaluation-harness.

Launching server

We launch the vLLM server following vLLM Recipes using container image vllm/vllm-openai-rocm:v0.31.0.

VLLM_ROCM_USE_AITER=1 \
VLLM_ROCM_USE_AITER_FP8BMM=0 \
VLLM_ROCM_USE_AITER_FP4BMM=0 \
vllm serve amd/GLM-5.3-Quark-MXFP4-AttnFP8 \
  --served-model-name GLM-5.3-MXFP4 \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 8 \
  --kv-cache-dtype fp8_e4m3 \
  --trust-remote-code \
  --max-model-len 400000 \
  --max-num-seqs 128 \
  --reasoning-parser glm47 \
  --tool-call-parser glm47 \
  --enable-auto-tool-choice

Note: GPQA Diamond, AIME26, and MMLU Pro were evaluated with --max-model-len 327680; SciCode requires --max-model-len 400000.

Evaluating model in a new terminal

GSM8K

pip install lm-eval[api]

export VLLM_ROCM_USE_AITER=1
export VLLM_ROCM_USE_AITER_FP8BMM=0
export VLLM_ROCM_USE_AITER_FP4BMM=0

lm_eval --model vllm \
  --model_args "pretrained=amd/GLM-5.3-Quark-MXFP4,tensor_parallel_size=4,dtype=auto,quantization=quark,max_model_len=32768,trust_remote_code=True,max_num_seqs=32" \
  --tasks gsm8k \
  --batch_size auto

GPQA Diamond

evalscope eval \
  --model GLM-5.3-MXFP4 \
  --api-url http://127.0.0.1:8000/v1/chat/completions \
  --datasets gpqa_diamond \
  --eval-batch-size 128 \
  --repeats 8 \
  --generation-config '{"temperature":1.0,"top_p":0.95,"max_tokens":320000,"timeout":43200}' \
  --work-dir logs/gpqa_diamond

AIME26

evalscope eval \
  --model GLM-5.3-MXFP4 \
  --api-url http://127.0.0.1:8000/v1/chat/completions \
  --datasets aime26 \
  --eval-batch-size 128 \
  --repeats 16 \
  --generation-config '{"temperature":1.0,"top_p":0.95,"max_tokens":320000,"timeout":7200}' \
  --work-dir logs/aime26

MMLU Pro

evalscope eval \
  --model GLM-5.3-MXFP4 \
  --api-url http://127.0.0.1:8000/v1/chat/completions \
  --datasets mmlu_pro \
  --eval-batch-size 128 \
  --repeats 1 \
  --generation-config '{"temperature":1.0,"top_p":0.95,"max_tokens":320000,"timeout":43200}' \
  --work-dir logs/mmlu_pro

SciCode

evalscope eval \
  --model GLM-5.3-MXFP4 \
  --api-url http://127.0.0.1:8000/v1/chat/completions \
  --datasets scicode \
  --eval-batch-size 128 \
  --repeats 8 \
  --generation-config '{"temperature":1.0,"top_p":0.95,"max_tokens":320000,"timeout":43200,"extra_body":{"chat_template_kwargs":{"clear_thinking":true}}}' \
  --dataset-args '{"scicode":{"review_timeout":1800,"extra_params":{"provide_background":true}}}' \
  --sandbox '{"enabled":true,"engine":"docker","pool_size":16}' \
  --work-dir logs/scicode

AA-LCR (v1.1)

Judge configuration (aa-lcr-judge.json; replace <YOUR_SUBSCRIPTION_KEY>):

{
  "strategy": "llm",
  "models": [
    {
      "model_id": "gpt-5.6-luna",
      "api_url": "https://llm-api.amd.com/openai/responses",
      "api_key": "EMPTY",
      "eval_type": "openai_responses_api",
      "generation_config": {
        "reasoning_effort": "medium",
        "max_tokens": 4096,
        "timeout": 7200,
        "extra_query": {
          "api-version": "2025-04-01-preview"
        },
        "extra_headers": {
          "Ocp-Apim-Subscription-Key": "<YOUR_SUBSCRIPTION_KEY>"
        }
      }
    }
  ]
}

evalscope eval \
  --model GLM-5.3-MXFP4 \
  --api-url http://127.0.0.1:8000/v1/chat/completions \
  --datasets aa_lcr \
  --dataset-hub huggingface \
  --eval-batch-size 64 \
  --repeats 16 \
  --generation-config '{"temperature":1.0,"top_p":0.95,"max_tokens":16384,"timeout":43200}' \
  --dataset-args '{"aa_lcr":{"hf_dataset_revision":"9a77ef56b717057ade24ceab4d273712a0b4f19e"}}' \
  --judge "$(cat aa-lcr-judge.json)" \
  --work-dir logs/aa_lcr

IFBench

evalscope eval \
  --model GLM-5.3-MXFP4 \
  --api-url http://127.0.0.1:8000/v1/chat/completions \
  --datasets ifbench \
  --eval-batch-size 128 \
  --repeats 5 \
  --generation-config '{"temperature":1.0,"top_p":0.95,"max_tokens":320000,"timeout":43200}' \
  --dataset-args '{"ifbench":{"primary_metric":"prompt_level_loose"}}' \
  --work-dir logs/ifbench

License

Downloads last month: 7,477

Safetensors

Model size

384B params

Tensor type

F32

BF16

F8_E4M3

Model tree for amd/GLM-5.3-Quark-MXFP4-AttnFP8

Base model

zai-org/GLM-5.3

Quantized

(118)

this model

Evaluation results

harborframework/terminal-bench-2.1 · Terminalbench 2 1 View evaluation results source leaderboard

88.2
datacurve/deep-swe · Deep Swe View evaluation results source leaderboard

66.9
cais/hle · Hle View evaluation results source

62.5 ^*
harborframework/terminal-bench-3.0 · Terminalbench 3 View evaluation results source

28.3