SenseTime has introduced SenseNova Loop-9B-Preview, a multimodal model available in the open source. This model utilizes a recomputation method to increase processing depth without needing to duplicate the corresponding model weights. The announcement of this release was made on October 30th.
The model is capable of accepting both text and image data as input, generating text to perform tasks related to document understanding, code writing, and external tool usage.
It functions by running the common block of 24 decoder layers twice, which allows for the creation of a deeper computational path based on a base model with approximately nine billion parameters.
The documentation training configuration supports sequences up to 262,144 tokens and integrates examples including text, images, code, and tool usage. SenseTime has published the model weights, inference instructions, and a training guide, distributing the project under the Apache 2.0 license.