Instructions to use ludovicoYIN/MiniMax-M2-BF16-W4A16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ludovicoYIN/MiniMax-M2-BF16-W4A16 with Transformers:

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

pipe = pipeline("text-generation", model="ludovicoYIN/MiniMax-M2-BF16-W4A16", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ludovicoYIN/MiniMax-M2-BF16-W4A16", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("ludovicoYIN/MiniMax-M2-BF16-W4A16", trust_remote_code=True, 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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ludovicoYIN/MiniMax-M2-BF16-W4A16 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ludovicoYIN/MiniMax-M2-BF16-W4A16"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ludovicoYIN/MiniMax-M2-BF16-W4A16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ludovicoYIN/MiniMax-M2-BF16-W4A16

SGLang

How to use ludovicoYIN/MiniMax-M2-BF16-W4A16 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 "ludovicoYIN/MiniMax-M2-BF16-W4A16" \
    --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": "ludovicoYIN/MiniMax-M2-BF16-W4A16",
		"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 "ludovicoYIN/MiniMax-M2-BF16-W4A16" \
        --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": "ludovicoYIN/MiniMax-M2-BF16-W4A16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ludovicoYIN/MiniMax-M2-BF16-W4A16 with Docker Model Runner:
```
docker model run hf.co/ludovicoYIN/MiniMax-M2-BF16-W4A16
```

MiniMax-M2-BF16-W4A16

This repository contains a quantized checkpoint produced with llm-compressor from the base model MiniMaxAI/MiniMax-M2.

What this model is

Base model: MiniMaxAI/MiniMax-M2
Quantization pipeline: llm-compressor
Quantization recipe: AWQModifier
Scheme: W4A16
Main quantized targets: MoE expert MLP weights (w1, w2, w3)
MoE gates and lm_head are excluded from quantization per recipe

How it was generated

This model was generated in the llm-compressor workspace using the MiniMax M2 quantization flow in examples/quantizing_moe/minimax_m2_example.py.

Reproduction steps:

Prepare environment and install llm-compressor dependencies.
Set model_id in examples/quantizing_moe/minimax_m2_example.py to the BF16 base checkpoint path.
Run the example script:
- python examples/quantizing_moe/minimax_m2_example.py
The script applies AWQModifier with W4A16 on MiniMax M2 MoE experts (w1/w2/w3) and saves the compressed checkpoint.
Output directory is created as:
- MiniMax-M2-BF16-W4A16

Notes

This is a derived quantized artifact, not an official upstream release from MiniMaxAI.
Inference quality/performance may differ from the original BF16 checkpoint depending on workload and hardware.

Downloads last month: 11

Safetensors

Model size

34B params

Tensor type

I64

I32

BF16

Model tree for ludovicoYIN/MiniMax-M2-BF16-W4A16

Base model

MiniMaxAI/MiniMax-M2

Quantized

(44)

this model