SenseNova U1.5

SenseTime's NEO-unify architecture drops the bolted-together design typical of multimodal models: SenseNova U1.5 processes vision and language end to end without a separate visual encoder or VAE, so one checkpoint handles visual question answering, native 4K text-to-image generation, multi-image editing, and interleaved text-image output. The flagship U1.5-8B-MoT pairs roughly 8B understanding parameters with 8B generation parameters, and editing accepts bounding-box and reference-image controls for precise regional changes across several input images. An H100 renders a 2048px image in about 9 seconds; with Q4 quantization and balanced offloading the model squeezes into 10 to 12 GB of VRAM, though 24 GB or more is recommended, and 8-step LoRA variants trade quality for speed. The August 2026 release is Apache 2.0 across GitHub, Hugging Face, and ModelScope, an unusually permissive grant for a unified generation model of this strength, following the original U1 in April and a U1.5 preview in July, with around 5.5k GitHub stars.