SenseTime drops encoder-free NEO-unify multimodal model

The "encoder-free" trend is hitting its stride, proving that raw pixel processing can rival specialized VAEs with significantly higher data efficiency.

• –Eliminates CLIP/SigLIP dependencies, reducing architectural bloat and lowering inference latency for local edge deployments.
• –Mixture-of-Transformer backbone allows simultaneous visual understanding and image generation without the performance trade-offs of modular systems.
• –Flow matching on raw pixels achieves 31.56 PSNR, nearly matching Flux's VAE while remaining a completely unified architecture.
• –Extreme data efficiency—outperforming counterparts like Bagel with fewer tokens—suggests a more scalable path for native multimodal pre-training.