Nucleus Image: A New Open-Source Sparse MoE Text-to-Image Model - quasa.io

Yes, another text-to-image model just dropped — and this one is genuinely interesting.

Meet Nucleus Image, the latest release from Nucleus AI, a small San Francisco-based AI company founded in 2023. While it’s still very early, the model brings something fresh to the open-source image generation scene: a sparse Mixture-of-Experts (MoE) Diffusion Transformer.

Key Highlights

• 17B total parameters, but only ~2B active per forward pass (thanks to 64 routed experts + 1 shared expert).
• They call it “the 1st Sparse MoE Diffusion Transformer”.
• 32-layer architecture, with 29 layers using sparse MoE instead of dense FFN (first 3 layers remain dense for training stability).
• Uses Grouped-Query Attention.
• Text encoder: Qwen3-VL-8B-Instruct.
• VAE: Qwen-Image VAE (16-channel).

The model is released as a base model — no DPO, RLHF, or heavy human preference tuning yet. According to the team, this is intentional: they want to release a strong foundation first.

---

Training Scale

What makes it particularly noteworthy is the dataset size:

• ~1.5 billion image-text training pairs;
• ~700 million unique images.

That’s serious scale for an open-source effort.

---

Current Status (as of May 2026)

• Weights are available;
• Detailed technical report and model card published;
• Claimed “Day 0” support in Hugging Face Diffusers;
• Code has not been released yet (despite heavy “truly open” messaging).

Faces are still a bit weak (typical for early base models), but overall image quality looks promising for its efficiency class.

The model is small enough that it should (with some effort) fit into **16GB VRAM**, and the sparse MoE design suggests fast inference relative to its total parameter count.

Also read:

• Bixonimania: The Fake Disease That Fooled Every Major AI — And Then Sneaked Into a Real Medical Journal
• The Mirage Effect: Stanford Just Proved That “Computer Vision” Is Often Just Confident Bullshit
• The Billion-Dollar Solo Startup: How One Man Built Medvi — $401M in Year One, $1.8B Run Rate in Year Two — With Just AI and His Brother

---

Why This Matters

Sparse MoE architectures have already revolutionized language models (Mixtral, DeepSeek, etc.). Bringing the same efficiency breakthrough to diffusion models could be a big deal — especially if the code drops and the community starts fine-tuning it.

Nucleus AI is a small team (appears to be 2–10 people based on public profiles), yet they’re swinging big: previously releasing a 22B-token 500B LLM back in 2023. This feels like a serious attempt to compete in the image generation space.

We’ll have a better idea of real-world performance once the full code and inference scripts are public. Until then, you can check out samples on their site:

→ https://withnucleus.ai/image

Another day, another impressive open-source drop. The image generation race is far from over — and the gap between closed and open models continues to shrink.