SenseNova-Loop-9B: Open model runs a 24-layer block twice with shared weights

What's going on

On October 30, 2026, the SenseNova team released the model SenseNova-Loop-9B-Preview on Hugging Face. It is a multimodal language model that reads text and images and outputs text. It is aimed at long contexts and agent workflows, meaning tasks where a model reads documents, writes code and drives external tools. The weights are available under the Apache 2.0 license, and the code is on GitHub at OpenSenseNova/SenseNova-Loop. It is explicitly a preview release.

What's special: a block that runs twice

The usual rule is that more compute depth means more layers and therefore more parameters. SenseNova-Loop takes a different route and uses a "loop decoder." Here is the structure according to the Model Card:

• Layers 0 to 3 process the input once.
• Layers 4 to 27 (24 layers) form a block that is applied twice in a row with the same weights. The source writes this as B(B(h)).
• Layers 28 to 31 and the output head produce the result.

That gives the model 32 unique layers, but it executes 56 layers per forward pass. Instead of about 9B parameters, the effective compute size comes to roughly 14.6B ("inference parameters"). The key point: on the second pass, the block doesn't receive the same input as on the first. It receives the already transformed intermediate state. The test author on note.com compares this to applying the same tool again to a half-finished workpiece.

This is complemented by mHC (manifold-constrained Hyper-Connections) in a two-stream variant. Two residual streams run inside the loop section (a residual stream is the side path that carries information past layers). Learned mappings control how information flows between the two streams.

More architecture and training details

• Hybrid attention: The decoder alternates in groups of four between linear attention (cheaper for long texts) and full attention.
• Context: The longest training sequences span 262,144 tokens. Training combined text, images, code and tool trajectories.
• Format: The model is exported in BF16 for the main body and FP32 for the mHC routing parameters.
• Demos: The Model Card shows agent examples such as dynamically designed videos, presentations (for example on Teachers' Day and an Iceland trip), data analysis of a lunar mission simulation, and energy consumption analyses. The model itself only outputs text, though. It does not generate video directly, as the note.com author stresses.

What the double execution does not mean

The author on note.com checked the published code and warns against some misunderstandings:

• The loop is not a self-test of the finished app. The model doesn't open a browser or click buttons to find bugs. The repeated computation happens only inside the layers.
• The number of passes is fixed. According to the author, it doesn't adapt dynamically to the difficulty of the request.
• Shared weights save memory for weights, but not necessarily compute time or runtime memory: in the code, each pass gets its own cache (a buffer that stores earlier computations).

If you build apps with it, you should still check whether the code meets your requirements and whether the model can fix errors after you feed back real test results.

What you need to try it

The sources give no concrete hardware figures. The test author neither loaded the weights nor ran inference, so memory needs, speed, Japanese quality and the reproducibility of the examples are all unverified. As a rough estimate, 4-bit quantization should need about 6–8 GB of VRAM (video memory), while BF16 needs considerably more (around 18 GB for the weights alone). The KV cache grows because of the double execution and long contexts. The only officially documented route is a pinned Transformers environment with custom code or an SGLang version with special modifications. It's therefore unclear whether common tools such as Ollama or llama.cpp already work. Most likely they don't yet.

Our take

Gaining compute depth by reusing layers rather than adding parameters is interesting for locally run models: smaller weight files with higher effective performance. But the evidence isn't there yet. There is no technical report, and the available sources include no benchmarks or comparisons with similarly sized models. Compute costs rise with the 56 layer executions. Only the weight memory stays small. As a preview with special-purpose code, it's more for developers and tinkerers than for quick everyday use. SenseNova is SenseTime's model family. Earlier versions such as SenseNova 5.0 and V6 were large MoE models that weren't open. The sources don't make clear whether Loop-9B is related to them.