Sparse MoE efficiency
According to the model card, 17B total capacity activates only about 2B parameters per forward pass through 64 routed experts per MoE layer plus one shared expert.
Open Source Model Profile · NucleusAI
Nucleus-Image
Nucleus-Image is a text-to-image MoE diffusion model from NucleusAI. According to the model card, it pairs 17B total parameters with about 2B active per pass and is released as a base model.
- Publisher
- NucleusAI
- Task
- text-to-image
- Model type
- Unknown
- License
- apache-2.0
- Library
- diffusers
- Publication status
- Approved for indexing
Model overview
Nucleus-Image is published by NucleusAI as a text-to-image model. Captured Safetensors metadata reports about 16.9B parameters under apache-2.0 with diffusers support. According to the model card, it is a 32-layer sparse MoE diffusion transformer with 17B total parameters, about 2B active per pass, released as a base model without post-training optimization.
Recorded capabilities
Expert-Choice routing
According to the model card, routing uses Expert-Choice with a decoupled design separating timestep-aware assignment from timestep-conditioned computation.
Text KV caching
According to the model card, text tokens contribute only as key-value pairs and their projections cache across denoising steps via TextKVCacheConfig in diffusers.
Multi-aspect training
According to the model card, aspect-ratio bucketing runs from the outset with progressive 256 to 512 to 1024 resolution training and seven documented output sizes.
Use cases in the source record
- Text-to-image generation for commercial, conceptual, and everyday imagery using the documented diffusers pipeline.
- Multi-aspect-ratio image experiments using the card's seven documented sizes and text KV-caching setup.
Limitations and unknowns
- No context-window value was extracted from this record.
- No independent evaluation results were extracted; published GenEval, DPG-Bench, and OneIG-Bench figures are publisher claims.
- Provider state is historical snapshot data, not independently refreshed current availability.
- Quality-versus-efficiency and comparison statements about other image models are publisher claims and were not independently verified.
Source and provenance
Source: NucleusAI/Nucleus-Image
Captured: Unknown. Processed: 2026-09-07T19:34:35.117136+00:00.
🌐 Website | 🖥️ GitHub | 🤗 Hugging Face | 📑 Tech Report Introduction Nucleus-Image is a text-to-image generation model built on a sparse mixture-of-experts (MoE) diffusion transformer architecture. It scales to 17B total parameters across 64 routed experts per layer while activating only ~2B parameters per forward pass, establishing a new Pareto frontier in quality-versus-efficiency. Nucleus-Image matches or exceeds leading models including Qwen-Image, GPT Image 1, Seedream 3.0, and Imagen4 on GenEval, DPG-Bench, and OneIG-Bench. This is a base model released without any post-training optimization (no DPO, no reinforcement learni…
F001F002F003F004F005F006F008F009F010F011F012F013F014F015F016F017F018F019