SenseTime Open-Sources Loop-9B: Same Layers, Run Twice

For years the scaling race has pointed one way: more parameters, more layers. SenseTime's newly open-sourced SenseNova Loop-9B-Preview asks a different question. If the parameter count stays fixed, can the model think twice?

The answer is in the architecture. Loop-9B's decoder has only 32 distinct layers, but the block from layer 4 through 27 runs twice in sequence with the same weights, the second pass taking the first pass's output as input. Each forward pass therefore executes 56 decoder layers while storing just 32 sets of weights. The model has about 9B parameters in total and roughly 14.6B in effective inference terms. SenseTime sums it up as scaling compute depth, not model size.

Looping alone is not new. The hard part is that running one block twice tends to muddle information as it goes. SenseTime's fix is two-stream mHC, manifold-constrained hyper-connections: inside the looped section the model keeps two residual streams and uses learned, Sinkhorn-normalized mappings to route and merge information between them. Research this year, including Hyperloop Transformers, showed that adding hyper-connections to a looped middle block can match or beat depth-matched standard models with about half the parameters. Loop-9B brings that line of work to a usable multimodal size.

Its positioning is clear as well. It takes text and images and outputs text. Attention follows a hybrid pattern of three linear-attention layers and one full-attention layer per four-layer group, and training sequences reach about 262K tokens. Official examples cover long, agentic jobs such as motion design, slide generation and data analysis. SenseTime publishes comparisons against peer models, but the Preview tag is a reminder that this is an interim release.

The significance goes beyond one 9B model. On-device and private deployments are usually limited by memory rather than compute, and looping trades time for space: spend more inference compute, save a large chunk of weight storage. For phone, PC and edge vendors, that is more useful than yet another bigger model. The costs are real too: higher latency at the same parameter count, and serving frameworks still need to catch up on shared-weight loops.

Most open releases from Chinese model makers this year have competed on data and post-training over familiar architectures. This time SenseTime is betting on the architecture itself. Whether the Preview grows into a full release, and whether loop depth can keep climbing, is worth watching.