Model Architecture: MiniMaxM2ForCausalLM
- Input: Text
- Output: Text
Model Optimizations:
- Weight quantization: FP4
Intended Use Cases:
- Reasoning.
- Function calling.
- Subject matter experts via fine-tuning.
- Multilingual instruction following.
- Translation.
Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws).
Release Date: 03/28/2026
Version: 1.0
Model Developers: RedHat (Neural Magic)

Model Optimizations

This model was obtained by quantizing the weights and activations of MiniMax-M2.5 to FP4 data type. This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 75%. Only the weights and activations of the linear operators within transformers blocks of the language model are quantized.

Deployment

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

python

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "RedHatAI/MiniMax-M2.5-NVFP4"
number_gpus = 1
sampling_params = SamplingParams(temperature=1.0, top_p=0.95, top_k=40, min_p=0, max_tokens=256)

messages = [
    {"role": "user", "content": prompt}
]

tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [{"role": "user", "content": "Give me a short introduction to large language model."}]

prompts = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

llm = LLM(model=model_id, tensor_parallel_size=number_gpus)

outputs = llm.generate(prompts, sampling_params)

generated_text = outputs[0].outputs[0].text
print(generated_text)

vLLM aslo supports OpenAI-compatible serving. See the documentation for more details.

Creation

Creation details

This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below.

python

import torch
from datasets import load_dataset
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer

from llmcompressor import oneshot
from llmcompressor.modeling.minimax_m2 import (  # noqa: F401
    CalibrationMiniMaxM2SparseMoeBlock,
)
from llmcompressor.modifiers.quantization import QuantizationModifier

# Load the model
model_id = "RedHatAI/MiniMax-M2.5-BF16"
config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, dtype=torch.bfloat16, config=config,trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
# MoE calibration is handled automatically by the pipeline.
# The `CalibrationMiniMaxM2SparseMoeBlock` modules (from
# `llmcompressor.modeling.minimax_m2`) will be applied during calibration to enable
# proper expert calibration. These replace the original
# `MiniMaxM2SparseMoeBlock` class from
# `transformers.models.minimax_m2.modeling_minimax_m2`.

# Select calibration dataset.
DATASET_ID = "HuggingFaceH4/ultrachat_200k"
DATASET_SPLIT = "train_sft"

# Select number of samples. 512 samples is a good place to start.
# Increasing the number of samples can improve accuracy.
NUM_CALIBRATION_SAMPLES = 512
MAX_SEQUENCE_LENGTH = 2048

# Load dataset and preprocess.
ds = load_dataset(DATASET_ID, split=f"{DATASET_SPLIT}[:{NUM_CALIBRATION_SAMPLES}]")
ds = ds.shuffle(seed=42)


def preprocess(example):
    return {
        "text": tokenizer.apply_chat_template(
            example["messages"],
            tokenize=False,
        )
    }


ds = ds.map(preprocess)


# Tokenize inputs.
def tokenize(sample):
    return tokenizer(
        sample["text"],
        padding=False,
        max_length=MAX_SEQUENCE_LENGTH,
        truncation=True,
        add_special_tokens=False,
    )


ds = ds.map(tokenize, remove_columns=ds.column_names)

moe_ignores = [
    "lm_head",
    "re:.*block_sparse_moe.gate$",
]

# Experts live under `model.layers.*.block_sparse_moe.experts.<idx>.(w1|w2|w3)`.
EXPERT_TARGET_REGEX = [
    "re:.*block_sparse_moe\\.experts\\.\\d+\\.w1$",
    "re:.*block_sparse_moe\\.experts\\.\\d+\\.w2$",
    "re:.*block_sparse_moe\\.experts\\.\\d+\\.w3$",
]

recipe = QuantizationModifier(
    targets=EXPERT_TARGET_REGEX,
    scheme="NVFP4",
    weight_observer="mse",
    ignore= moe_ignores
)


# Apply algorithms.
oneshot(
    model=model,
    dataset=ds,
    processor=tokenizer,
    recipe=recipe,
    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
    max_seq_length=MAX_SEQUENCE_LENGTH,
    sequential_targets=["MiniMaxM2DecoderLayer"],
)

# Save to disk compressed.
SAVE_DIR = model_id.rstrip("/").split("/")[-1] + "-NVFP4"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)

Evaluation

The model was evaluated on the ifeval, mmlu_pro and gsm8k_platinum using lm-evaluation-harness, on reasoning tasks using lighteval. vLLM was used for all evaluations.

Evaluation details

Deploy using vllm to create an OpenAI-compatible API endpoint:

vLLM:

vllm serve RedHatAI/MiniMax-M2.5-NVFP4 --max-model-len 262144 --reasoning-parser deepseek_r1

lm-evaluation-harness

lm_eval --model local-chat-completions \
  --tasks mmlu_pro_chat \
  --model_args "model=RedHatAI/MiniMax-M2.5-NVFP4,max_length=262144,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=64,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=1200" \
  --num_fewshot 0 \
  --apply_chat_template \
  --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,top_k=40,min_p=0.0,max_gen_toks=64000

lm_eval --model local-chat-completions \
  --tasks ifeval \
  --model_args "model=RedHatAI/MiniMax-M2.5-NVFP4,max_length=262144,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=64,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=1200" \
  --num_fewshot 0 \
  --apply_chat_template \
  --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,top_k=40,min_p=0.0,max_gen_toks=64000

lm_eval --model local-chat-completions \
  --tasks gsm8k_platinum_cot_llama \
  --model_args "model=RedHatAI/MiniMax-M2.5-NVFP4,max_length=262144,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=64,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=1200" \
  --num_fewshot 0 \
  --apply_chat_template \
  --gen_kwargs "do_sample=True,temperature=1.0,top_p=0.95,top_k=40,min_p=0.0,max_gen_toks=64000

lighteval

lighteval_model_arguments.yaml

yaml

model_parameters:
  model_name: RedHatAI/MiniMax-M2.5-NVFP4
  dtype: auto
  gpu_memory_utilization: 0.9
  max_model_length: 40960
  generation_parameters:
    temperature: 1.0
    top_k: 40
    min_p: 0.0
    top_p: 0.95
    max_new_tokens: 64000

lighteval endpoint litellm lighteval_model_arguments.yaml  \
  "aime25|0,math_500|0,gpqa:diamond|0"

Accuracy

Benchmark	RedHatAI/MiniMax-M2.5-BF16	RedHatAI/MiniMax-M2.5-NVFP4	Recovery (%)
GSM8k Platinum (0-shot)	95.15	93.91	98.70
IfEval (0-shot)	92.05	89.89	97.66
AIME 2025	87.50	77.08	88.10
GPQA diamond	83.67	80.30	95.98
Math 500	87.33	87.73	100.46
MMLU Pro Chat	80.83	80.08	99.07