SenseTime has released SenseNova-Loop-9B, an open-weight multimodal model that runs part of its network twice per forward pass to get more computation out of a small parameter budget. The model card on Hugging Face lists the release date as November 17, with AWQ and GGUF quantized weights added on November 18. It follows a preview version released on October 31.
The model has 32 distinct layers. According to SenseTime, layers 0 to 3 and 28 to 31 run once, while the middle block of layers 4 to 27 runs twice, for 56 layer executions per forward pass. SenseTime puts the model at about 9 billion total parameters and about 14.6 billion inference-equivalent parameters. It uses hybrid attention, with three linear-attention layers followed by one full-attention layer in each group of four, plus two-stream manifold-constrained hyper-connections.
The context window is 512K tokens, and the weights are released under the Apache 2.0 license. SenseTime says that with 4-bit weights and an 8-bit KV cache, multimodal inference at a 256K context needs about 15 GiB of GPU memory, with the 4-bit weights taking about 4.9 GB.
On SenseTime's reported benchmarks, the model scores 59.55 on Terminal-Bench 2.1 and 69.90 on GAIA, up from 40.45 and 37.86 for the preview. It roughly matches Qwen3.8-27B on AIDABench (65.31 vs 65.01) and AutomationBench (36.67 each), but trails it on NL2Repo (41.36 vs 44.60). The model runs two loops by default, and SenseTime reports that some tasks improve with three or four. It lists long-horizon agent work, coding and complex instruction following as weaknesses.
For developers, the release is a test of whether looped layers can give smaller models more capability on a single consumer-class GPU. The 15 GiB figure at 256K context puts long-context multimodal inference within reach of a 16 GB card.