SenseNova-Loop-9B-Preview Released on HuggingFace for Long Context and Agentic Tasks

On HuggingFace, on October 28, 2026, the model sensenova/SenseNova-Loop-9B-Preview was published with the text-generation task. It is a multimodal language model that accepts text and images and generates text. The author positions it for long context and agentic scenarios: document parsing, coding, interaction with external tools. The maximum training sequence length is 262,144 tokens.

What it means

The model contains 9.7 billion parameters; the weights in the repository occupy 35.5 GB in safetensors format. The license is apache-2.0. It runs via transformers; the architecture is autoregressive. There are already five quantized builds on the site — GGUF and NVFP4. The most compact working version is Q4_0 from guideboardlabs, weighing 5 GB. The main feature of SenseNova-Loop-9B-Preview is a looped decoder: a block of 24 layers is executed twice with shared weights, which yields 56 layer executions with 32 physical layers. This increases processing depth without doubling the parameters of the repeated block. Dual-stream hyper-connections with manifold constraints govern the transfer of information between the two residual streams inside the looped section. For local running, 5 GB is enough on the quantized Q4_0 build, which is noticeably less than the full weight volume. More on local AI tool news — in the CPU3D news feed.