Nucleus-Image is an open-source sparse mixture-of-experts (MoE) diffusion model designed for efficient image generation and manipulation. Released in 2026, the model represents a significant advancement in parameter-efficient deep learning architectures for generative tasks, combining large model capacity with computational efficiency through sparse activation patterns.

Nucleus-Image comprises 17 billion total parameters with only 2 billion parameters actively engaged during inference, achieving a computational efficiency ratio of approximately 8.5:1 between total and active parameters. This architecture leverages sparse mixture-of-experts techniques, which allow the model to selectively activate different expert subnetworks based on input characteristics, thereby reducing computational overhead while maintaining expressive capacity ¹⁾.

The model operates as a diffusion-based generative system, building upon the diffusion probabilistic models framework established in recent generative AI research ²⁾.

Nucleus-Image was released with comprehensive support for the open-source Hugging Face Diffusers library from day-0, enabling immediate integration into existing machine learning pipelines and applications. This compatibility represents a significant practical advantage for practitioners seeking to incorporate the model into production systems.

The complete release package includes training code and dataset recipes, distributed under the Apache 2.0 open-source license. This transparency enables researchers and practitioners to reproduce training results, fine-tune the model for domain-specific applications, and contribute improvements to the codebase. The provision of training recipes allows downstream users to understand the data preparation, preprocessing, and training procedures used to develop the model ³⁾.

The sparse MoE architecture employed by Nucleus-Image addresses a fundamental challenge in scaling deep learning models: the computational cost and memory requirements increase substantially with model size. By implementing conditional computation—where only a subset of parameters activate for each input—the model achieves significant efficiency gains ⁴⁾.

Sparse MoE systems typically implement gating mechanisms that route inputs to specific expert subnetworks based on learned routing functions. This approach maintains the theoretical capacity of a large model while substantially reducing the computational footprint during inference and training. The 2 billion active parameters suggest that Nucleus-Image utilizes selective expert routing, potentially based on input-dependent gating networks.

As a diffusion-based generative model, Nucleus-Image is applicable to various image generation and manipulation tasks, including:

* Text-to-image synthesis and conditional image generation * Image inpainting and content modification * Style transfer and artistic image manipulation * Data augmentation for computer vision applications * Creative and design-focused applications

The efficient parameter utilization makes deployment feasible on resource-constrained hardware, extending accessibility to practitioners with limited computational budgets.

The immediate compatibility with the Hugging Face Diffusers library positions Nucleus-Image within a well-established ecosystem for diffusion models. The Diffusers library provides standardized interfaces for model loading, inference, and fine-tuning, reducing implementation complexity for downstream users ⁵⁾.

The Apache 2.0 license and release of training code align with best practices in open-source machine learning research, promoting transparency, reproducibility, and collaborative development. The provision of dataset recipes enables the scientific community to validate training procedures and understand data requirements for comparable sparse MoE diffusion models.

• Sparse Mixture of Experts Diffusion
• Sparse Mixture of Experts (MoE)
• Mixture-of-Experts (MoE) Architecture