Model release · SenseTime
SenseNova-Loop-9B-Preview
- Released
- Review updated
SenseTime's preview multimodal model for long-context and agent work. Instead of adding parameters, it runs a middle block of layers twice with the same weights to get deeper compute, and uses manifold-constrained hyper-connections (mHC) to move information between layers. SenseTime says it beats Qwen3.5-27B on WorkBuddy Bench, RaccoonPPT Bench and AutomationBench. It's aimed at coding, office documents and agent tasks, and was trained on contexts of up to about 256K tokens.
Availability
Open weights under Apache 2.0, downloadable from Hugging Face with code on GitHub. The model card covers running it with Transformers or LightLLM, and MTP speeds up inference. A technical report has not been published yet.