Instructions to use amd/MiniMax-M3.1-Flash-MXFP4-QADv1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use amd/MiniMax-M3.1-Flash-MXFP4-QADv1 with Transformers:

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

pipe = pipeline("image-text-to-text", model="amd/MiniMax-M3.1-Flash-MXFP4-QADv1", trust_remote_code=True)
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("amd/MiniMax-M3.1-Flash-MXFP4-QADv1", trust_remote_code=True)
model = AutoModelForMultimodalLM.from_pretrained("amd/MiniMax-M3.1-Flash-MXFP4-QADv1", trust_remote_code=True, device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use amd/MiniMax-M3.1-Flash-MXFP4-QADv1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "amd/MiniMax-M3.1-Flash-MXFP4-QADv1"
# 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/MiniMax-M3.1-Flash-MXFP4-QADv1",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/amd/MiniMax-M3.1-Flash-MXFP4-QADv1

SGLang

How to use amd/MiniMax-M3.1-Flash-MXFP4-QADv1 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/MiniMax-M3.1-Flash-MXFP4-QADv1" \
    --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/MiniMax-M3.1-Flash-MXFP4-QADv1",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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/MiniMax-M3.1-Flash-MXFP4-QADv1" \
        --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/MiniMax-M3.1-Flash-MXFP4-QADv1",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use amd/MiniMax-M3.1-Flash-MXFP4-QADv1 with Docker Model Runner:
```
docker model run hf.co/amd/MiniMax-M3.1-Flash-MXFP4-QADv1
```

MiniMax-M3.1-Flash MXFP4 QADv1

Pure-MXFP4 MiniMax-M3.1-Flash checkpoint for AMD Instinct MI350/MI355. Drop-in replacement for amd/MiniMax-M3.1-Flash-Quark-MXFP4.

Model Overview

Model Architecture: MiniMaxM3SparseForConditionalGeneration
- Input: Text, Image, Video
- Output: Text
Supported Hardware Microarchitecture: AMD MI350/MI355
Inference Engine: vLLM (same build and flags as amd/MiniMax-M3.1-Flash-Quark-MXFP4)
Model Optimizer: AMD-Quark + GPTQ rounding + quantization-aware distillation (QAD)
- Routed-expert weight quantization: OCP MXFP4, Static (pure MXFP4, no mixed precision)
- Routed-expert activation quantization: OCP MXFP4, Dynamic
- All other layers: same precision/format as amd/MiniMax-M3.1-Flash-Quark-MXFP4, with QAD-tuned values

Config, quantization config, tokenizer, processor files, safetensors shard layout, tensor names/dtypes/shapes, and the dspark draft model match the AMD MXFP4 release. Only weight values differ.

Method

Pure MXFP4 routed experts (OCP MX, E2M1 values, E8M0 block-32 scales), same format and scales as amd/MiniMax-M3.1-Flash-Quark-MXFP4.
GPTQ rounding of the routed-expert FP4 codes.
QAD: on-policy distillation of non-expert weights from the official MiniMax NVFP4 release (MiniMaxAI/MiniMax-M3.1-Flash), including agentic coding trajectories. Experts stay frozen on the GPTQ MXFP4 grid.

Evaluation (DeepSWE)

Same vLLM ROCm A4W4 setup for every model (customer a4w4_reproduce recipe). NVFP4 was run in emulation on MI350. DeepSWE: 113 tasks, 2 runs.

Benchmark	Plain MXFP4	NVFP4	QADv1 (this model)	QADv1 vs plain MXFP4 [95% CI]
DeepSWE, whole benchmark (113 tasks, 2 runs)	58.4%	60.2%	61.5%	+3.1 [−4.0, +10.2]

Notes

Point estimate is above plain MXFP4 and at/slightly above NVFP4; not statistically significant at this benchmark size.

Model serving

Serve exactly like amd/MiniMax-M3.1-Flash-Quark-MXFP4; only the model path changes.

vLLM

VLLM_ROCM_USE_AITER=1 \
vllm serve /path/to/MiniMax-M3.1-Flash-MXFP4-QADv1 \
    --tensor-parallel-size 4 \
    --reasoning-parser minimax_m3 \
    --block-size 128 \
    --speculative-config '{"method":"dspark","model":"/path/to/MiniMax-M3.1-Flash-MXFP4-QADv1/dspark","num_speculative_tokens":7,"attention_backend":"FLASHINFER_TS","kv_cache_dtype":"fp8","draft_sample_method":"probabilistic","rejection_sample_method":"block"}' \
    --disable-hybrid-kv-cache-manager

Evaluation used "attention_backend":"TRITON_ATTN" with "num_speculative_tokens":3; outputs are expected to be identical.

Integrity

SHA256SUMS lists the sha256 of every file. Check with sha256sum -c SHA256SUMS.

License

Same as the base model; see LICENSE. Credits: MiniMax, AMD Quark / MXFP4 release.

Downloads last month: -

Safetensors

Model size

220B params

Tensor type

F32

BF16

F8_E4M3

Model tree for amd/MiniMax-M3.1-Flash-MXFP4-QADv1

Base model

MiniMaxAI/MiniMax-M3.1-Flash

Quantized

(40)

this model