SenseTime releases SenseNova U1 models on HuggingFace

SenseTime open-sourced two SenseNova U1 multimodal models on HuggingFace with open weights: an 8B dense MoT model and a ~3B-activated MoT model.

The models use the NEO-Unify architecture, which replaces separate visual encoders and VAEs with a unified token space for multimodal understanding, reasoning, and generation.

Released under Apache 2.0, the models are commercially usable and available for developers to build on.

They target tasks like VQA, interleaved text-image generation, and content creation such as PPTs, posters, and diagrams.