5h agoMainland ChinaAI Large Language ModelsAI Open Source ModelsAI Data Analytics & SpreadsheetsAI Programming
SenseTime launches SenseNova Loop-9B with 27B-level performance on low-memory hardware
SenseTime has launched the SenseNova Loop-9B model, continuing its work on parameter efficiency. The model reuses layer parameters through a Loop structure. It scales compute depth at about 9B parameters for complex reasoning, coding, tool use and multimodal tasks.
The model turns 32 physical decoder layers into 56 layer calculations. It accepts text and images and outputs text, with up to 512K tokens of context. The official 4-bit version weighs about 4.9GB and can run within 10GB of video memory.
In tests, it scored 41.36 on NL2Repo, close to Qwen3.8-27B's 44.60. It scored 65.31 on AIDABench, above the other model's 65.01. The weights and code are open, and an AWQ 4-bit version was available on launch day.
Comments
Loading comments…