SenseNova-Loop-9B: SenseTime's open 9B model that runs its layers twice
Open-weight 9B multimodal model that reuses its middle layers twice (32 physical layers, 56 layer executions) with 512K context under Apache 2.0, matching Qwen3.8-27B on agentic benchmarks.
Why it matters
Gets 27B-class agentic results from 9B weights via a looped architecture — a genuinely different efficiency trick — with official AWQ and GGUF builds from day one.
What you could build with it
A small team could run a 9B model that behaves like a 27B one on a single GPU for agentic coding and data-analysis tasks, keeping inference in-house instead of paying frontier API prices.
Does it hold up?
Too early to judge: released Nov 17 with strong vendor benchmarks (GAIA 69.90, Terminal-Bench 2.1 59.55), but no independent deployment reports yet; SGLang and vLLM support landed the same week, so serving it is at least straightforward.
Built with SenseNova-Loop-9B
- SenseNova-Loop-9B runs 24 of its layers twice to act like a bigger modelarticle · Explainer of the looped architecture (written for the Oct 31 Preview, same architecture as the full release): 56 layer computations from 32 stored layers, and the compute cost of a 14.6B model.
- SenseNova-Loop-9B runs its middle layers twice, openlyarticle · Coverage of the open release: 9B multimodal model under Apache 2.0, 27B-class results claimed on agentic tasks, technical report still pending.
- sgl-project/sglanggithub · 39,024 ★ · SGLang added SenseNova Loop support (PR #47281) so the model can be served with its inference engine.
- vllm-project/vllmgithub · 97,154 ★ · vLLM added SenseNova Loop support (PR #64832) for serving the model at scale.
Learn more
First spotted on hf: source.
More AI models
Edge0-145B: open 145B MoE that runs on a single consumer GPUEdge0 open-sourced 145B and 120B latent-MoE models derived from Qwen3.8-Flash-Next, plus an inference framework that…models · JEV 0.74Strata: run a 125B model on a gaming PCFree, MIT-licensed local inference engine that runs Qwen3.8-Flash-Next (125B MoE, about 6B active per token) on…models · JEV 0.73Qwen4-Max: Alibaba's new flagship model goes liveAlibaba's 2.4T-parameter flagship is now generally available via API, with a 1M-token context, native vision, and…models · JEV 0.67smolmaxxing: run 16B LLMs on a 1 GB TV box or a 2013 iPadAn MIT-licensed llama.cpp patch set that runs language models up to a 16B MoE on devices with about 1 GB of usable RAM…models · JEV 0.65Blue TTS v3: a 388MB open text-to-speech model with voice cloningAn open-source 388MB text-to-speech model (MIT) with voice cloning from about 10 seconds of speech, 7 languages, 15…models · JEV 0.65Haru-4BOpen full-duplex spoken-dialogue foundation model that listens and speaks simultaneously (7.5 Hz frames, 133 ms per…models · JEV 0.63
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.