ChinaSenseTime released a model that runs 24 layers twice.
The repeat deepens computation without duplicating weights, so the model stays near 9 billion parameters.
The shared block loops twice, and the model reads text and images to produce text for documents, coding and tool use.
SenseTime published the weights under Apache 2.0, so outside labs can now test it.
How each outlet framed it
- TechNode
- explains the looping mechanism: a 24-layer block run twice, giving depth without duplicated weights, under Apache 2.0
Sources: TechNode