SenseTime releases open-source AI model using layer reuse technology

SenseTime unveiled its SenseNova Loop-9B-Preview model on October 30, introducing a technique that reuses model layers to simulate deeper processing. By running a shared block of 24 decoder layers twice, the model achieves greater complexity from a base of approximately nine billion parameters without the overhead of additional weights.

The multimodal model processes both text and image inputs, targeting document analysis, software development, and tool-use applications. It supports context windows of up to 262,144 tokens, a capacity that typically requires significantly larger hardware configurations in traditional model architectures.

The company has published the weights, inference instructions, and training details under the Apache 2.0 license. This move allows developers to inspect and integrate the technology into their own stacks, providing an alternative to monolithic model designs.