Fine-Tune MiniMax-M3

Introduction

MiniMaxAI/MiniMax-M3 is MiniMaxAI’s 428B A22B Mixture-of-Experts first vision-language model combining long-context reasoning, agentic workflows, and creative capabilities in a single platform. The multimodal MoE model evolved from MiniMax M2 along three axes — width, attention, and visual grounding.

MiniMax-M3 enables advanced use cases such as long-form video understanding, extended coding tasks (8+ hours), and high-quality design workflows.

To set up your environment to fine-tune this model with NeMo AutoModel, follow the installation guide.

Model Overview

Architecture

• Model type: 428B total / 22B active MoE vision-language model.
• Language module: MiniMax-M2.7 backbone with 60 layers (3 dense + 57 MoE), 64 attention heads, 128 experts, block-sparse DSA attention on the MoE layers and a 512k context length.
• Vision module: CLIP-style ViT with 32 layers and dynamic resolution image input from 336×336 up to 2016×2016
• Precision targets: BF16 and MXFP8
• Hardware target: trained on Hopper GPUs.

Data

Multimodal Supervised Fine-Tuning Data

Use image/video instruction data that matches the target agent workflow. Good candidates include:

• frontend mockup-to-project examples,
• screenshot-debugging conversations,
• structured data-processing tasks with visual context,
• image/video question-answer pairs for bounded task execution.

For a full walkthrough of how multimodal datasets are preprocessed and integrated into NeMo AutoModel, including chat-template conversion and collate functions, see the Multi-Modal Dataset Guide.

Launch Training

NeMo AutoModel supports several ways to launch training: the AutoModel CLI with Slurm, interactive sessions, torchrun, and more. For full details on Slurm batch jobs, multi-node configuration, and environment variables, see the Run on a Cluster guide.

Standalone Slurm Skeleton

Before running, make sure your cluster environment is configured following the Run on a Cluster guide.

$
export TRANSFORMERS_OFFLINE=1
$
export HF_HOME=/path/to/hf_cache
$
export HF_DATASETS_OFFLINE=1
$
export WANDB_API_KEY=your_wandb_key
$
$
srun --output=output.out \
>
     --error=output.err \
>
     --container-image /path/to/automodel26.04.image.sqsh \
>
     --no-container-mount-home bash -c "
>
  CUDA_DEVICE_MAX_CONNECTIONS=1 automodel \
>
  /path/to/minimax_m3_vl.yaml \
>
  --nproc-per-node=8 \
>
  --model.pretrained_model_name_or_path=/path/to/MiniMax-M3 \
>
  --processor.pretrained_model_name_or_path=/path/to/MiniMax-M3"

Full fine-tuning recipe can be found at: examples/vlm_finetune/minimax_m3 /minimax_m3_vl_sft_ep32pp4.yaml and LoRA recipe at: examples/vlm_finetune/minimax_m3 /minimax_m3_vl_lora_pp4ep8_8node.yaml

Before you start:

• Clone or mirror the model checkpoint locally before launching a multi-node run.
• Ensure HF_HOME points to a shared cache visible from all nodes.
• Cache the dataset locally if running with HF_DATASETS_OFFLINE=1.
• Configure the wandb section in the recipe to record loss, throughput, and memory curves.

Training Results

The SFT and LoRA fine-tuning loss curves are shown below.

SFT

LoRA