SenseTime officially released SenseNova Loop-9B. This model continues the Preview version's exploration of parameter efficiency: by reusing model layer parameters through the Loop structure and combining it with manifold-constrained hyper-connections (mHC), it expands computational depth at approximately 9B parameters, targeting complex reasoning, code, tool calling, and multimodal tasks.
Benefiting from the efficient loop structure, the official 4-bit quantized weights are only about 4.9GB, with deployment memory requirements as low as under 10GB, significantly broadening the range of deployable hardware and enabling low-memory chips to achieve performance comparable to cutting-edge 27B-level models.
Since the Preview release, we have received community feedback and attention regarding model usage, inference deployment, and ecosystem adaptation. This official version retains the Loop and dual-stream mHC architecture, further optimizing the model's performance in Agent, code, and data analysis tasks, while continuing to explore the role of recurrent computation in enhancing model capabilities.
SenseNova Loop-9B has approximately 9B total parameters. By executing the shared intermediate layers twice, the model expands 32 physical decoder layers into 56 layer computations; it supports text and image input with text output, and a context length of up to 512K tokens.
In Agent and data analysis evaluations, the model demonstrates competitiveness comparable to Qwen3.8-27B: an NL2Repo score of 41.36, close to the latter's 44.60; and an AIDABench score of 65.31, higher than the latter's 65.01.
Model Highlights
Shared Recurrent Module: Layers 4–27 form a shared-weight recurrent module executed twice, enabling the model to complete 56 decoder layer computations in a single forward pass.
Dual-Stream mHC: Information routing and fusion of two residual streams in the recurrent module via learnable mappings.
Hybrid Attention: Decoders are grouped in sets of four layers, using a configuration of three layers of linear attention and one layer of full attention.
Long-Context Multimodal: Training covers text, images, code, and tool call trajectories, with context length extended up to 512K tokens.
Low-Resource Private Deployment Friendly: With 4-bit weight quantization and 8-bit KV cache, memory requirements for 256K long-context inference are approximately 15 GiB, helping lower the barrier for local and private deployment.
Official Quantized Version Available Day 0: The official AWQ 4-bit quantized version is provided on the first day of release for quick deployment.
SenseNova Loop-9B has released model weights and related code for developers and researchers to download, deploy, and explore.
Model Capabilities: Official Version Further Improved Over Preview
We evaluated the official SenseNova Loop-9B on tasks including general Agent, code, and data analysis, comparing it with the Preview version and models like Qwen3.8-27B.
Evaluation results show that the official SenseNova Loop-9B has significantly improved over the Preview version on multiple benchmarks; at approximately 9B parameters, it demonstrates performance on par with Qwen3.8-27B.
Model Performance