All news

SenseTime open-sources SenseNova Loop-9B weights: 9 billion parameters and 512K-token context

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

The model with a 256 thousand token context fits in roughly 15 GiB of video memory, so long documents and agentic tasks can run on a single graphics card.

On 17 November SenseTime officially released the SenseNova Loop-9B model and opened its weights and code, simultaneously posting an official 4-bit AWQ variant on release day. It is a multimodal decoder with roughly 9 billion total parameters (about 14.6 billion "equivalent" at inference, per the model card) and a context of up to 512K tokens. The architecture is a Looped Transformer: the block of layers from the 4th to the 28th is executed twice with shared weights, so 32 physical layers yield 56 layer computations per pass. Attention alternates (three linear layers to one full one), and two residual mHC "streams" are added; text and images go in, text comes out. With 4-bit weights (about 4.9 GB) and an 8-bit KV cache, inference at a 256K context requires roughly 15 GiB of video memory. Per the model card, the release beats the preview version on all seven benchmarks: AutomationBench 14.33 → 36.67, Terminal-Bench 2.1 40.45 → 59.55, GAIA 37.86 → 69.90. On release day, Chinese accelerator makers MetaX and Cambricon announced support — the so-called Day-0 adaptation.