AI/TLDR

SenseTime · 2026-05-12 · notable

SenseNova-U1 — 8B Dense and 30B-A3B MoE Native Unified Multimodal Models With NEO-Unify Pixel-Space Backbone, Apache 2.0

SenseTime's open-source unified multimodal family drops the vision encoder + VAE split, jointly trains understanding and generation on raw pixels, and hits 0.91 on GenEval. Code and weights on GitHub and HF.

SenseNova-U1 GitHub repository banner

An open native-unified multimodal model that ditches the vision encoder and VAE: one transformer reads and draws pixels end-to-end.

Key specs

GitHub stars1,593
Dense params8B
Moe total params30B
Moe active params3B
Geneval0.91
Hf papers trending#1

What is it?

SenseNova-U1 is SenseTime's open multimodal model family with two variants: an 8B dense (SenseNova-U1-8B-MoT) and a 30B-total / 3B-active Mixture-of-Experts (SenseNova-U1-A3B-MoT). Unlike most VLMs, it has no pretrained vision encoder and no VAE — it operates directly on pixels and text using a Mixture-of-Transformers (MoT) where understanding and generation streams share routing but keep decoupled parameters.

How does it work?

The NEO-unify backbone uses near-lossless 32×32 patch encoding/decoding to move pixels in and out without a learned tokenizer. Understanding is autoregressive over text; generation is pixel-space flow matching. The two objectives are co-trained in a single transformer so the model can perceive, reason, and generate within one pass — useful for interleaved tasks like image editing, text-rich infographics, and vision-language-action.

Why does it matter?

Unified models have been the long-promised replacement for the vision-encoder + LLM + diffusion-decoder Frankenstein stack. SenseNova-U1 is one of the first openly licensed (Apache 2.0) takes that posts competitive numbers across MMMU, MMBench, and GenEval (0.91), with code, weights, GGUF builds, and a hosted demo. The paper is the #1 trending paper on Hugging Face today.

Who is it for?

ML researchers and infra teams evaluating unified multimodal architectures.

Try it

https://huggingface.co/collections/sensenova/sensenova-u1

Sources · 4 outlets

Tags

  • multimodal
  • vision-language
  • unified-model
  • moe
  • mixture-of-transformers
  • neo-unify
  • open-weights
  • apache-2

← All releases · Learn AI