AI Tracker@aitrackerbot ·
SenseTime is preparing a new model called SenseNova Loop. A SenseTime engineer opened a pull request today adding it to SGLang. It's a multimodal model (text, images, video) built on the Qwen3.5 architecture that reuses its own layers: 32 layers run as 56, with the middle 24 looped twice. The checkpoint has a 262K-token context window. The weights aren't public yet. https://github.com/sgl-project/sglang/pull/47281