9B-parameter looped multimodal model from SenseTime that reuses a layer block twice for long-context and agentic tasks.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
SenseNova-Loop-9B-Preview is a multimodal language model from SenseTime, aimed at long-context and agentic workflows. It accepts text and images and produces text, for uses such as document understanding, coding and tool interaction. The release is labelled a preview.
The architecture uses a looped decoder with two-stream manifold-constrained hyper-connections (mHC). A block of 24 decoder layers (layers 4–27) is run twice with shared weights, giving 56 layer executions from 32 distinct layers. This means about 9B stored parameters but roughly 14.6B inference parameters. The decoder also interleaves linear and full attention in a repeating four-layer pattern (three linear, one full). Training recipes cover sequences up to 262,144 tokens, with text, images, code and tool trajectories.
Weights and code are published under Apache-2.0 on GitHub and loaded through Hugging Face Transformers with a custom implementation (README cites Transformers 5.18.0 and PyTorch 2.11.0 with CUDA 13.0). MTP inference acceleration is available; Dspark and Dflash2 are still in training. The checkpoint's loop configuration must be reproduced by the inference implementation, so compatible runtimes are required.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
9B looped multimodal model that repeats layers for more computation without more parameters. Competitive with larger models on agentic benchmarks; Apache 2, about 38GB.
Closest neighbours by meaning, not just by category.
Model · SenseTime
Open-weights 8B image generation and editing model from SenseTime that works in pixel space without a VAE.
Model · SenseTime
Open-weights 27B model from SenseTime that creates short videos by generating and rendering executable code.
Model · Google
Google's open-weights 12B multimodal (text and image input) instruction-tuned model from the Gemma 3 family.
Model · Recursive Intelligence
Dense 65B open-weight language model from Ricursive Intelligence, trained for agentic coding with up to 1M context.
Model · Google
Google's open-weights 12B encoder-free multimodal model in the Gemma 4 family, taking text, image and audio input.
Model · Black Forest Labs
9B-parameter open-weights image generation and editing model from Black Forest Labs, built for low-latency use.
