SenseTime releases open 9B AI model that runs its layers twice

SenseTime has released SenseNova-Loop-9B-Preview, an open-source model that reads text and images and writes text, with the announcement made on October 30. Its defining feature is architectural: instead of stacking more layers, it runs a shared block of 24 decoder layers twice, giving a deeper computation path from a model of roughly nine billion parameters.

The weights sit in a public Hugging Face repository under the sensenova organisation, and SenseTime has published inference instructions and training guidance alongside them, under the Apache 2.0 licence. What the sources do not contain is just as important for builders: no benchmark scores, from SenseTime or anyone else, appear in the material reviewed for this article.

What the Loop design does

Most language and vision-language models are built as a long stack of decoder layers, the repeated processing blocks that transform the input step by step before the model predicts its next token. A deeper stack usually means more parameters, more memory and larger files to download.

SenseNova-Loop-9B-Preview takes a different route. TechNode reports that it “uses repeated computation to extend its processing depth without duplicating the corresponding model weights”. Concretely, a shared block of 24 decoder layers is run twice, so data passes through the same weights a second time. The result, per the report, is a deeper computation path while the parameter count stays at approximately nine billion.

The practical appeal of this idea is straightforward. The amount of memory needed to hold a model’s weights depends on how many parameters it has, not on how many times those parameters are used. A looped design aims to get some of the benefit of depth while keeping the weight footprint of a 9B model. Whether that trade pays off in answer quality, and what it costs in speed, are questions the sources do not answer with numbers.

Specs at a glance

Item	What the sources state
Maker	SenseTime
Release name	SenseNova-Loop-9B-Preview
Parameters	Approximately nine billion
Architecture	Shared block of 24 decoder layers, run twice
Inputs	Text and images
Output	Text
Training sequence length	Up to 262,144 tokens in the documented training setup
Training data mix	Text, images, code and tool-use examples
Licence	Apache 2.0
Published materials	Weights, inference instructions, training guidance

What the model is built to do

SenseTime aims the model at three kinds of work: document understanding, coding, and tasks that involve external tools. That focus matches the training mix TechNode describes, which combines text, images, code and tool-use examples.

The chat template shipped in the Hugging Face repository shows how those capabilities are wired in practice:

• Image inputs are inserted into a conversation as vision placeholder tokens, and the template can optionally number them (“Picture 1”, “Picture 2”) so a prompt can refer to specific images.
• Tool calling uses an XML-style format: the model is told to reply with a <tool_call> block containing a <function=...> element and named <parameter=...> entries. Tool results come back to the model wrapped in <tool_response> tags.
• Reasoning text is separated from the final answer with <think> tags, and the template keeps that reasoning only for turns after the latest user query.
• System messages cannot contain images or videos; the template raises an error if they do.

The same template appears in the repository snapshot at commit 3462613, which gives developers a fixed version to pin when testing.

One detail is worth flagging. The template includes handling for a video placeholder as well as images. TechNode, however, describes the model as accepting text and images. A template branch is not the same as a supported, tested capability, so video input should be treated as unconfirmed until SenseTime documents it.

Benchmarks: no scores in the sources yet

This is the part of the story builders will look for first, and the sources reviewed here do not provide it. Neither the Hugging Face material available to this desk nor TechNode’s report includes benchmark results, whether from SenseTime’s own evaluation or from an independent leaderboard.

That matters for two reasons. First, the central claim about the Loop design is that reusing layers yields “deeper processing”. Without comparisons against a conventional 9B model, or against SenseTime’s own non-looped models, there is no evidence in the sources that the extra depth translates into better answers. Second, the “Preview” label signals that SenseTime itself does not present this as a finished release. Any numbers that do appear later should be read with attention to who ran them and what they were compared against.

Until then, the honest position is that the architecture is documented and the weights are available, but performance is unknown.

Access and licence

The model is distributed through Hugging Face under the sensenova/SenseNova-Loop-9B-Preview repository. SenseTime has published the weights together with inference instructions and training guidance, which means teams can run the model themselves and, in principle, fine-tune it.

The licence is Apache 2.0, a permissive licence widely used in open-source software that allows commercial use and modification. That stands out against some recent open releases with narrower terms, such as Baidu’s 280B ERNIE base model, whose licence excludes the EU, UK and Korea. The sources do not mention a hosted API from SenseTime for this model, regional restrictions, or a waitlist.

What it means for builders

For developers working on document pipelines, code assistants or agents that call tools, a 9B open model with image input and an Apache 2.0 licence is a practical size: small enough to self-host on modest hardware compared with frontier-scale models, and licensed for commercial products. It joins a run of open multimodal releases aimed at practical tasks, including Kakao’s open model that reads text and images to make decisions.

The limits and open questions are clear from what the sources leave out:

• Quality. No benchmark results are available in the sources, so there is no basis yet for comparing it with other 9B-class vision-language models.
• Speed and cost. Running a 24-layer block twice keeps the weights small, but the sources give no figures on latency, throughput or memory use during inference. Teams should measure this on their own hardware before planning around it.
• Context length. The 262,144-token figure describes the documented training setup. The sources do not state how well the model performs on very long inputs in practice.
• Tool-call format. The XML-style tool-call format differs from the JSON-style function calling many APIs use, so agent frameworks may need a parser adapter.
• Preview status. The name itself marks this as an early release, and behaviour or files may change in later versions.

What to watch next

The next useful signals are concrete. First, published evaluation results: either SenseTime’s own scores with clear comparisons, or third-party testing on public leaderboards, would show whether the Loop design delivers more than a standard 9B model. Second, inference measurements from developers running the model, which will reveal what the second pass through the shared layers costs in speed. Third, clarification on video input, given that the chat template includes a video path while the described inputs are text and images. Finally, whether SenseTime follows the preview with a full release or larger looped variants will indicate how central this design is to its open model plans.