A Guth Labs publication ↗
Back to News

SenseTime adds quantized download links for SenseNova-Loop-9B

Brushed metal processor atop two green glass layers on a cream pedestal.
AI-generated editorial illustration

SenseTime added official AWQ and GGUF download links for SenseNova-Loop-9B on November 18, according to the model’s public repository. The update follows the full model’s November 17 release and an earlier preview on October 31.

The roughly nine-billion-parameter model accepts text and images and generates text. Its architecture reuses a block of layers: 32 distinct decoder layers perform 56 layer computations, increasing computational depth without storing a separate set of weights for every step. The repository lists a context window of up to 512,000 tokens.

Downloads include four-bit AWQ weights and several GGUF precision options. SenseTime estimates approximately 15 GiB of GPU memory for a 256,000-token context with four-bit weights and an eight-bit key-value cache. That is a configuration-specific estimate, not a universal hardware requirement.

The project uses Apache 2.0 licensing. Its documentation acknowledges inconsistent coding and agent performance, including difficulty following complex instructions.

The additional download options give developers more ways to test a compact multimodal model locally, with memory requirements and compatibility still dependent on their chosen setup.

Sources and citations

Revision history

Published .