Nucleus AI (@withnucleusai) - Twitter Profili | Zamantika Mersobahis Locabet

Sabitlenmiş Tweet

Nucleus AI@withnucleusai·17h

Introducing Nucleus-Image: the first sparse Mixture-of-Experts diffusion model 17B parameters. Only 2B active. 10x more parameter-efficient than leading diffusion models. Toe-to-toe with GPT Image 1, Imagen 4, and Qwen-Image: from pure pre-training alone. No DPO. No RL. No preference tuning. Day 0 support in 🤗 Hugging Face diffusers. Fully open-source under Apache 2.0. Weights, training code, and dataset recipe - we're not holding anything back <3

English

26

74

533

35.1K

Nucleus AI@withnucleusai·5h

@pwnies The only right way to do it.

English

0

7

66

Jacob Miller@pwnies·12h

> Weights, training code, and dataset recipe Holy moly. Peeps don’t realize how invaluable this is. One of the first true open source drops I’ve seen.

Nucleus AI@withnucleusai

Introducing Nucleus-Image: the first sparse Mixture-of-Experts diffusion model 17B parameters. Only 2B active. 10x more parameter-efficient than leading diffusion models. Toe-to-toe with GPT Image 1, Imagen 4, and Qwen-Image: from pure pre-training alone. No DPO. No RL. No preference tuning. Day 0 support in 🤗 Hugging Face diffusers. Fully open-source under Apache 2.0. Weights, training code, and dataset recipe - we're not holding anything back <3

English

0

3

12

1.4K

Nucleus AI@withnucleusai·17h

Get started in 3 lines of code with native 🤗 diffusers integration including Text KV caching across all denoising steps out of the box. 🤗 Model: huggingface.co/NucleusAI/Nucl… 🖥️ GitHub: github.com/WithNucleusAI/… 📑 Tech Paper: arxiv.org/abs/2604.12163 🌐 Website: withnucleus.ai/image 📚 Blog: huggingface.co/blog/NucleusAI…

English

1

2

30

1.9K

Nucleus AI@withnucleusai·17h

Introducing Nucleus-Image: the first sparse Mixture-of-Experts diffusion model 17B parameters. Only 2B active. 10x more parameter-efficient than leading diffusion models. Toe-to-toe with GPT Image 1, Imagen 4, and Qwen-Image: from pure pre-training alone. No DPO. No RL. No preference tuning. Day 0 support in 🤗 Hugging Face diffusers. Fully open-source under Apache 2.0. Weights, training code, and dataset recipe - we're not holding anything back <3

English

26

74

533

35.1K

Nucleus AI@withnucleusai·17h

What's under the hood: → Decoupled MoE routing: fixes silent expert collapse in DiTs → Text tokens as KV-only: full caching across all 50 denoising steps → Muon optimizer + checkpoint averaging: no EMA needed → Trained on 700M images, 1.5B captions, curriculum-sampled → Post-training, higher-resolution variants, and controllable generation to be unlocked next.

English

0

1

18

1.8K

Nucleus AI@withnucleusai·17h

The benchmarks speak for themselves: DPG-Bench: 88.79 (#1 overall) GenEval: 0.87 (matches Qwen-Image beats GPT Image1) OneIG-Bench: 0.522 (beats Imagen 4 & Recraft V3) All from a base model. Imagine what post-training will unlock!!

English

4

0

24

2.1K

Nucleus AI

Keşfet