Minimax-h3-Turbo-SLA

MiniMax-H3 Turbo-SLA

MiniMax-H3 Turbo-SLA is a 4-step distilled FL2V checkpoint for MiniMax-H3, enhanced with SLA (Sparse–Linear Attention) for more efficient inference.

The model uses an 85% attention sparsity ratio. In our tested LightX2V setup, it delivers approximately 2.5× inference acceleration on an NVIDIA RTX 5090 while retaining competitive visual quality.

This repository contains LoRA weights and requires the original MiniMax-H3 model for inference. Actual performance may vary with resolution, video length, software environment, and hardware configuration.

Demo

Side-by-side comparison between the 30-step MiniMax-H3 base model and the 4-step MiniMax-H3 Turbo-SLA model:

<video controls loop muted playsinline width="100%"> <source src="https://cdn-uploads.huggingface.co/production/uploads/680de13385293771bc57400b/HXV6GbED84PeLteFXqK5l.mp4" type="video/mp4"> Your browser does not support the video tag. </video>

Model Highlights

• 4-step distillation for substantially reduced denoising steps.
• SLA sparse attention with an 85% sparsity ratio.
• Approximately 2.5× inference acceleration on RTX 5090 in our tested LightX2V setup.
• 768p FL2V generation support.
• Native LightX2V and ComfyUI-compatible BF16 checkpoints.

Available Checkpoints

LightX2V Inference

For installation and general MiniMax-H3 inference instructions, refer to the LightX2V MiniMax-H3 examples.

Use the following configuration for SLA-enabled inference on an RTX 5090:

minimax_h3_fp8_4step_5090_vae_fp8_sla.json

Key SLA-related settings in this configuration include:

{
  "attn_type": "dynamic_sparse_attn",
  "dynamic_sparse_attn_setting": {
    "sparsity_ratio": 0.85,
    "operator": "sage2"
  },
  "video_flow_shift": 6.0,
  "audio_flow_shift": 3.0,
  "h3_step_update": "training_euler"
}

Use the configuration file linked above as the source of truth for the complete inference setup, including FP8 DiT/VAE options and checkpoint paths.

ComfyUI

For ComfyUI workflows, use:

minimax_h3_fl2v_turbo_4step_v0.1_768p_sla_comfyui_bf16.safetensors

The ComfyUI checkpoint is a converted version of the SLA-enabled 4-step LoRA and is intended for compatible MiniMax-H3 ComfyUI workflows.

Related Projects

• MiniMax-H3
• MiniMax-H3 Turbo
• LightX2V
• SLA: Sparse–Linear Attention

Acknowledgements

This work builds on MiniMax-H3, LightX2V, and SLA. We thank the respective authors and contributors for making their work available to the community.

Citation

If you use SLA in your work, please cite:

@article{zhang2025sla,
  title={SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse-Linear Attention},
  author={Zhang, Jintao and Wang, Haoxu and Jiang, Kai and Yang, Shuo and Zheng, Kaiwen and Xi, Haocheng and Wang, Ziteng and Zhu, Hongzhou and Zhao, Min and Stoica, Ion and others},
  journal={arXiv preprint arXiv:2509.24006},
  year={2025}
}

License

The adapter weights in this repository are released under the Apache 2.0 License. Use of the MiniMax-H3 base model is also subject to its corresponding license and terms of use.