SenseTime releases SenseNova-Loop-9B with 512K context
SenseTime's new 9B multimodal model reuses layers to reach 56 computations, supports up to 512K context and ships with an official 4-bit release.
SenseTime has released SenseNova-Loop-9B, a multimodal model with roughly 9 billion parameters, a stated context window of up to 512,000 tokens and an architecture that reuses shared layers to perform deeper computation.1
The weights and code are available to the community through the project's Hugging Face release. SenseNova-Loop-9B accepts text and images and generates text, positioning it for reasoning, coding, tool use and data-analysis workloads rather than as a text-only chat model.1
What makes the model different
The model uses 32 distinct decoder layers but repeatedly executes 24 middle layers, producing 56 layer computations in a forward pass. SenseTime pairs that looped design with two-stream manifold-constrained hyper-connections, or mHC, which route and mix information between residual streams. The result is an attempt to increase effective computation without scaling the parameter count in the usual way.1
The model card also describes hybrid attention, alternating linear and full attention in a repeating four-layer pattern. That is part of the explanation for the unusually long context claim: SenseTime says the training recipe includes text, images, code and tool-use trajectories, with the context length extended to 512K tokens.1
Deployment should be more approachable than the headline size suggests. SenseTime lists an official AWQ 4-bit release and says multimodal inference at a 256K context length needs about 15 GiB of GPU memory with 4-bit weights and an 8-bit KV cache.1 That is a vendor estimate, not a guarantee for every runtime or workload, and long-context performance will depend heavily on the serving stack.
The benchmark picture
SenseTime reports gains over its Preview release across the seven benchmarks shown in the project's comparison chart. The release scores 69.90 on GAIA, 59.55 on Terminal-Bench 2.1, 65.31 on AIDABench and 41.36 on NL2Repo.1 The model card compares some results with Qwen3.8-27B, where SenseNova-Loop-9B is ahead on AIDABench but behind on NL2Repo.1
Those numbers are developer-run results published by the project, not independent testing. They are useful for understanding what SenseTime is targeting, but they should not be treated as a settled ranking until outside researchers reproduce the evaluations under the same conditions.
Our opinion
SenseNova-Loop-9B is interesting because it attacks the cost of deeper reasoning with architectural reuse instead of simply adding more parameters. The 512K context claim is the attention-grabber, but the more practical detail is the official 4-bit path and the reported 15 GiB memory requirement at 256K context. If the model's long-context and multimodal behaviour survives independent testing, it could become a useful local option for developers who need more than a compact text model. For now, the release is a promising technical experiment backed by self-reported benchmarks, not a reason to rewrite the leaderboard.