---
license: mit
base_model: MiniMaxAI/MiniMax-M2.1
tags:
  - minimax
  - moe
  - nvfp4
  - quantized
  - vllm
  - blackwell
library_name: transformers
---

# MiniMax-M2.1-NVFP4

NVFP4 quantized version of [MiniMaxAI/MiniMax-M2.1](https://huggingface.co/MiniMaxAI/MiniMax-M2.1) for efficient inference on NVIDIA Blackwell GPUs.

## Model Details

| Property | Value |
|----------|-------|
| Base Model | [MiniMaxAI/MiniMax-M2.1](https://huggingface.co/MiniMaxAI/MiniMax-M2.1) |
| Architecture | Mixture of Experts (MoE) |
| Total Parameters | 229B |
| Active Parameters | ~45B (8 of 256 experts) |
| Quantization | NVFP4 (e2m1 format) |
| Size | 131 GB |

## Quantization Details

- **Format**: NVFP4 with two-level scaling (block-wise FP8 + global FP32)
- **Scheme**: `compressed-tensors` with `nvfp4-pack-quantized` format
- **Target**: All linear layers in attention and MoE experts
- **Group Size**: 16

## Requirements

- NVIDIA Blackwell GPU (RTX 5090, RTX PRO 6000, etc.)
- vLLM with flashinfer-cutlass NVFP4 support
- ~130 GB VRAM (TP=2 recommended for dual GPU setups)

## Usage with vLLM

```python
from vllm import LLM, SamplingParams

llm = LLM(
    model="GadflyII/MiniMax-M2.1-NVFP4",
    tensor_parallel_size=2,
    max_model_len=4096,
    gpu_memory_utilization=0.90,
    trust_remote_code=True,
)

sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.9,
    max_tokens=1024,
)

outputs = llm.generate(["Your prompt here"], sampling_params)
print(outputs[0].outputs[0].text)
```

## Performance

Tested on 2x RTX PRO 6000 Blackwell (98GB each):

| Prompt Tokens | Output Tokens | Throughput |
|---------------|---------------|------------|
| ~100 | 100 | ~73 tok/s |
| ~1260 | 1000 | ~72 tok/s |

## License

Same as base model - see [MiniMaxAI/MiniMax-M2.1](https://huggingface.co/MiniMaxAI/MiniMax-M2.1) for details.

## Acknowledgments

- [MiniMax](https://www.minimax.io/) for the original MiniMax-M2.1 model
- [vLLM](https://github.com/vllm-project/vllm) team for NVFP4 quantization support