SenseNova-Loop-9B Released: Looped Transformer Matches Larger Models
SenseNova has officially released SenseNova-Loop-9B under the Apache 2.0 license, showing massive performance gains over its preview version across key benchmarks like GAIA and AutomationBench. The model achieves scores competitive with much larger models like Qwen3.8-27B while utilizing an efficient 9B parameter footprint. This release demonstrates the viability of looped transformer architectures in significantly amplifying computational depth without scaling physical parameters, offering high performance on agentic and multimodal tasks within an efficient hardware footprint. It highlights how parameter-efficient designs can punch well above their weight class, making advanced reasoning more accessible for local deployment. The model features 32 physical layers where 24 middle layers are executed twice in sequence, resulting in 56 layer computations per forward pass alongside a two-stream manifold-constrained hyper-connection (mHC) design. With 4-bit weights and an 8-bit KV cache, multimodal inference requires approximately 15 GiB of GPU memory for a 256K context length.
## BACKGROUND
A standard transformer relies on a stack of distinct layers where every token passes through a fixed set of weights exactly once. In contrast, looped transformers repeatedly execute shared layers to allow latent reasoning and refine internal hidden states before generating text, effectively expanding computational depth within a compact parameter budget.