Taplio - Build a growth routine on LinkedIn
Mini toolsNucleus Image
By Nucleus AI
Nucleus-Image is a base text-to-image model that uses a 32-layer sparse mixture-of-experts diffusion transformer, with 29 layers using MoE blocks containing 64 routed experts plus 1 shared expert. The Hugging Face card says it has 17B total parameters, about 2B active parameters per pass, uses Qwen3-VL-8B-Instruct as the text encoder and Qwen-Image VAE as the image tokenizer, and was trained on 700M images with 1.5B caption pairs. It is released under Apache 2.0 with full weights, training code, and dataset, and is explicitly presented as a base model without DPO, RL, or human preference tuning.
New
Image
Gen 4
Overview
Nucleus-Image is Nucleus AI’s open-source text-to-image model built on a sparse MoE diffusion transformer. It scales to 17B total parameters with only about 2B active per forward pass, aiming for a strong quality-efficiency tradeoff, and the model card says it matches or exceeds leading systems on GenEval, DPG-Bench, and OneIG-Bench.
About Nucleus AI
Nucleus AI is an AI research company building general intelligence systems and large language models. The company has released open-source models on HuggingFace, including a 22B parameter LLM trained on 500B tokens of RefinedWeb, and Nucleus-X, a RetNet-based model offering faster inference and lower memory requirements.
Industry:
Artificial Intelligence
Company Size:
6
Location:
San Francisco, CA, US
Website:
withnucleus.ai
Tools using Nucleus Image
No tools found for this model yet.
Last updated: April 15, 2026
