SenseNova-Loop-9B-Preview

Model Nov 4, 2026

SenseNova's open-weight multimodal preview reuses a shared decoder block to perform 56 layer executions with 32 distinct layers.

Visit primary link 2 sources attached

SenseNova diagram shows a shared looped decoder block between pre-loop and post-loop layers, with a two-stream mHC detail below. — Author-supplied architecture diagram from the English model card. The preview applies its 24-layer middle block twice; the generic N-th pass notation in the diagram is not a claim that arbitrary extra passes improve quality. Source: SenseNova

multimodal self-hosted

What happened lately

SenseNova-Loop-9B-Preview timeline

Momentum Nov 4, 2026

A 9B model that runs its layers twice: SenseNova Loop's speed tradeoff

Community reports 3 sources

The model card dates the preview release to October 30, 2026. It accepts text and images and generates text. The authors describe about 9 billion stored parameters and an approximately 14.6-billion inference-parameter scale from repeated computation; those quantities do not imply fourteen billion unique weights or a measured speed advantage.

The Apache-2.0 card documents custom model code, up to 262,144-token training sequences and deployment paths requiring compatible implementations. A technical report remains forthcoming, and the provided examples are author demonstrations rather than independent tests.