Overview
SenseNova Loop-9B-Preview is SenseTime's open-weight take on a different route to capability: rather than growing parameters, run the same weights more than once. The model has 32 decoder layers, keeps the first and last four, and treats the middle 24 as a shared looping block executed twice, written as B(B(h)). One forward pass performs about 56 layer computations while storing 32 layers of weights, roughly 9B.
The hard part of looping is not the second pass but keeping it from scrambling the first. Loop-9B uses two-stream mHC connections inside the looping block, maintaining two residual streams with learned routing and mixing, constrained by a Sinkhorn normalisation iterated 20 times. Attention is arranged in groups of four, three linear layers plus one full attention layer, with the looping boundary aligned to a complete group.
The result is a training context of up to 262,144 tokens, text and image input, text output, aimed at long documents and agent workflows. It ships under Apache 2.0 on Hugging Face and GitHub, exported in BF16 with mHC routing parameters in FP32, and community GGUF quantisations exist, with Q8_0 around 8.9GB. SenseTime frames the 9B model as comparable to 27B-scale models; the technical report is still pending and the model is marked Preview.
Key Features
- Depth through looping: The middle 24 layers share weights and run twice, so a forward pass does about 56 layer computations while storing only 32 layers, trading memory for depth.
- Two-stream mHC: Two residual streams with learned routing and mixing, plus Sinkhorn-normalised mixing matrices, keep signal propagation stable across iterations.
- Hybrid attention: Each group of four decoder layers alternates three linear attention layers with one full attention layer, with loop boundaries aligned to full groups.
- 256K context: Training sequences reach 262,144 tokens, aimed at reading a full document in one pass and at multi-step agent tasks.
- Text and image input: An autoregressive multimodal decoder that accepts text and images and produces text, so charts and screenshots can go straight in.
- Open and commercially usable: Apache 2.0 weights load through Transformers, serve on SGLang or LightLLM, and run on 16GB devices through community quantisation.
Use Cases
- Feed a hundred-page contract or research report in one go and ask questions across sections.
- Drop it into an agent pipeline that needs multi-step reasoning over office tasks.
- Deploy privately on machines with limited VRAM, trading storage for task performance.
- Power long-document question answering over an internal knowledge base, including chart screenshots.
- Generate structured deliverables such as slide decks and data analyses, where SenseTime reports strong benchmark scores.
- Use it as a reproducible sample of a looping architecture to compare against dense models of similar size.
Pros
- 9B of stored weights keeps the VRAM floor low, which suits private deployments and small teams.
- Looped computation raises effective depth, and SenseTime's benchmarks put it ahead on agent and office tasks.
- A 262,144-token training context means long documents fit without chunking.
- Apache 2.0 licensing leaves room for commercial use and further training.
- Community GGUF builds run on 16GB devices.
- SGLang and LightLLM support make it easy to slot into existing inference services.
Pricing
Open source under Apache 2.0. Weights are published on Hugging Face as sensenova/SenseNova-Loop-9B-Preview and on GitHub in the OpenSenseNova/SenseNova-Loop repository, with a BF16 backbone and FP32 mHC routing parameters. Loading requires trust_remote_code to bring in the custom looping implementation, and the official examples use Transformers. Deployment notes mention SGLang and LightLLM, and community GGUF quantisations exist with Q8_0 around 8.9GB. A hosted API has no published price yet, so the practical route is running the weights yourself.
Summary
Loop-9B separates two things usually discussed together: how many parameters you store and how deep each token is computed. What it saves is weight storage and VRAM, not compute, since the 56 layer computations still happen, which makes it most interesting for private deployment on tight hardware rather than throughput-sensitive serving. SenseTime positions it against 27B models, but those numbers are self-reported and the technical report is not out, so the fairer reading today is a downloadable, reproducible sample of the looping-architecture route.
Version History
- SenseNova Loop-9B-Preview listed: 9B looping architecture open-sourced (2026-10-31): Form at listing: SenseTime's open-weight Looped Transformer multimodal model, with layers 4 to 27 sharing weights across two passes for about 56 layer computations per forward pass, roughly 9B stored and about 14.6B effective inference computation. Adds two-stream mHC with Sinkhorn-constrained mixing and grouped attention of three linear plus one full layer. Training context reaches 262,144 tokens with text and image input and text output. SenseTime reports scores above Qwen3.5-27B on BIRD-Dev, WorkBuddy Bench, AIDABench, TeamBench and RaccoonPPTBench, with mHC ablation gains of 4.42 and 13.34 points on GPQA Diamond and TeamBench. Apache 2.0 weights, community GGUF builds available.