Open-source SenseNova-U1 unifies image understanding and generation

不卷参数卷架构，这个开源模型把图像理解和生成统一了

SenseTime open-sourced two SenseNova-U1 models: an 8B version and a 38B-total MoE version using NEO-unify. The architecture removes VE and VAE, processes pixels directly, and generates 2048×2048 images in about 9 seconds on one H100/H200 node. The key item is interleaved text-image reasoning; 32K context, long-text rendering, and beta interleaved creation remain limits.

Why it matters: HKR-H/K/R all pass: the architecture hook is concrete, the post gives model sizes and latency, and open multimodal work matters to builders. It stays in 78–84 because it is not a top-tier general-model launch.