🧬 Nucleus-Image: 17B Open-Source Image Model That Rivals GPT Image 1 While Using Just 2B Active Parameters

What if a free, open-source image generator could match the quality of the best proprietary models — while using a fraction of the compute?

Nucleus AI just released **Nucleus-Image**, a text-to-image model built on a sparse Mixture-of-Experts (MoE) diffusion transformer. It packs 17 billion total parameters across 64 routed experts per layer, but only activates roughly 2 billion per forward pass. Think of it as having 64 specialist artists, but only calling on the right ones for each job.

The results speak for themselves: Nucleus-Image matches or outperforms GPT Image 1, Imagen4, Seedream 3.0, and Qwen-Image on major benchmarks including GenEval, DPG-Bench, and OneIG-Bench — and these are **base model results** with zero post-training optimization. No DPO, no RLHF, no human preference tuning. The ceiling is still untouched.

What makes this truly groundbreaking is the release scope. Nucleus AI published the full model weights, training code, and dataset under an **Apache 2.0 license** — making it the first fully open-source MoE diffusion model at this quality tier. Anyone can use it commercially, fine-tune it, or build products on top of it.

The architecture uses progressive resolution training (256 → 512 → 1024), multi-aspect-ratio support, and text KV caching for automatic inference speedup. It generates everything from fantasy art and product photography to typography and architecture.

The gap between proprietary and open-source image generation just got a lot smaller.