A proposal to inference software came first
On Oct. 29, an engineer at China's SenseTime proposed a change to the open-source inference software "SGLang," which runs AI models, to support a new model called "SenseNova Loop." @aitrackerbot, an X account that tracks the appearance of new models, reported it. The next day, Oct. 30, SenseTime released the weights of "SenseNova-6.8-Loop-9B-Preview" on Hugging Face.
Using the same layers twice
The model consists of 32 layers; after passing through the first 4 layers, it runs the middle 24 layers twice with the same weights and then passes through the last 4 layers. That means a single pass performs the computation of 56 layers. It has about 9 billion parameters, but inference requires computation equivalent to about 14.6 billion parameters. The architecture aims to increase the amount of thinking without adding weights.
It accepts text and images and responds in text. Training used sequences of up to 260,000 tokens. Every 4 layers combine three computationally light attention layers with one attention layer that looks at the whole input.
Performance shown only in charts
The company presented results on eight evaluations, including database queries, creating presentation slides and banking tasks, but the figures appear only in charts. The model is positioned as a preview, and some of the components for faster inference are still being trained. It is licensed under the permissive Apache 2.0 license.
