SenseTime open-sources 9-billion-parameter AI model that reuses layers for deeper processing

What happened

On October 30, SenseTime announced SenseNova Loop-9B-Preview. It is a multimodal model, which means it can take in both text and images. It writes text back for tasks like reading documents, writing code, and using outside tools. The company shared the model's weights (the learned numbers that make the model work), instructions for running it, and guidance for training it. The project uses the Apache 2.0 license, which lets others use and change it freely.

Key numbers and facts

• The model has about 9 billion parameters (the adjustable numbers inside an AI model).
• It runs a shared block of 24 decoder layers (the parts that build the output) two times.
• The training setup supports sequences of up to 262,144 tokens (small pieces of text or image data).
• Training data mixes text, images, code, and examples of tool use.

How the reuse trick works

Most AI models get deeper by adding new layers, and each new layer needs its own stored numbers. SenseTime instead runs the same layers again. This gives the model a longer path of thinking without making the file size bigger. Think of it like reading a paragraph twice to understand it better, instead of buying a longer book.

Why it matters

Open-source releases let students, researchers, and small teams try advanced AI for free. A model that gets more processing depth without more stored weights could run on less powerful hardware. It also shows one more way that AI builders are trying to get better results without making models huge.

What to watch

• How well the model performs on real tests for documents, code, and tool use.
• Whether other developers build on the released weights and training guidance.
• Whether the reuse-of-layers approach is adopted by other AI companies.