SenseNova Open-Sources Loop-9B, a 9B Looped-Decoder Multimodal Model With 512K Context and Apache License
Summary
SenseNova released SenseNova-Loop-9B under Apache 2.0 on Hugging Face, an autoregressive multimodal decoder that reuses 24 middle layers to run its 32 physical layers 56 times per forward pass (~14.6B inference-equivalent params). It supports 512K context, text+image input, and matches Qwen3.8-27B on AutomationBench (36.67) while beating it on AIDABench (65.31 vs 65.01). vLLM and SGLang support is landing via open PRs.
Originally reported by huggingface.co
Read the original article →Original headline: SenseNova Open-Sources Loop-9B, a 9B Looped-Decoder Multimodal Model With 512K Context and Apache License