November 2, 2026, (Inside AI) — SenseTime has open-sourced a multimodal model that gets more thinking depth without adding more parameters. The model, called SenseNova Loop-9B-Preview, reuses the same block of 24 decoder layers twice, running a shared set of weights through a longer computation path. The company announced the release on October 30.
The model accepts text and images and produces text. It targets document understanding, coding, and tasks that involve calling external tools. The weights, inference instructions, and training guidance are published under the Apache 2.0 license, which permits commercial use with minimal restrictions. The documented training setup supports sequences of up to 262,144 tokens and mixes text, images, code, and tool-use examples.
The significance lies in the economics of scale. Most labs push capability by enlarging parameter counts, which raises training and inference costs. SenseTime is testing a different lever: looping a fixed layer block to simulate depth while keeping the weight footprint at roughly nine billion parameters. If the approach holds up in independent tests, it offers a cheaper route to stronger reasoning and longer-context handling for companies that cannot afford trillion-parameter training runs.
Looping Weights Is Not New, But Open Release Is
Reusing layers is an old idea in neural network design. Early recurrent architectures and weight-sharing schemes in the 1990s explored similar logic. More recently, researchers have experimented with recurrent depth and looped transformers to trade compute for parameters. What is different here is the open release of a multimodal model at this scale with published training guidance.
Read: PrismML Releases Bonsai 2: Reasoning AI That Fits On Phones
SenseTime has not disclosed benchmark scores in the release materials. That omission matters. Loop-based models can behave differently from standard transformers. Repeated layers may improve certain reasoning chains but can also cause instability or diminishing returns. Without independent evaluation, claims about depth gains remain unverified. Inside AI could not independently verify the performance claims.
The 262,144-token context window is notable. Long-context handling has become a competitive front among open models. A nine-billion-parameter model with that window and multimodal input could appeal to developers building document analysis or agentic tool-use pipelines. The Apache 2.0 license lowers barriers for commercial deployment, a contrast to more restrictive open-weight licenses used by some competitors.
The release also fits a broader pattern. Chinese AI labs have increasingly published open-weight models to gain developer adoption and influence. SenseTime, known for its computer vision heritage, is positioning itself in the multimodal and tool-use space. The loop architecture could become a test case for whether compute efficiency can substitute for raw scale.
Analysts will watch two things. First, whether independent benchmarks confirm that looping 24 layers twice delivers measurable gains over a standard nine-billion-parameter model. Second, whether the training recipe generalizes. The published guidance includes tool-use examples, suggesting SenseTime expects agentic applications. If developers can reproduce the results, the approach may spread to other open models.
Read: Moonshot AI Releases Kimi K3 Architecture with 2.8T Parameters
The model is available now through SenseTime's release channels. Developers can access the weights and documentation under the Apache 2.0 terms. The company has not announced a hosted API or commercial support tier for this preview. For teams evaluating open multimodal models, the release adds a new option with an unusual architectural bet.
