SenseTime has released SenseNova-Loop-9B-Preview, an open-weight model that presents a unique challenge in parameter sizing. While it has approximately 9 billion parameters, its looped architecture requires 56 layer passes per token, making its inference parameters closer to 14.6 billion. This model requires specialized serving engines like llama.cpp, vLLM, or SGLang, as standard implementations do not yet support its looped structure. Testing revealed that CPU performance can be a bottleneck, and disabling the loop for faster processing leads to incorrect outputs, indicating the loop is integral to its trained functionality. AI
IMPACT Introduces a new model architecture that challenges traditional parameter sizing and requires specialized inference engines.
RANK_REASON Release of an open-weight model with a novel architectural feature. [lever_c_demoted from research: ic=1 ai=1.0]
- Apache Software License 2.0
- llama.cpp
- RTX 4000 Ada
- SenseNova-Loop-9B-Preview
- SenseTime
- SGLang
- transformers
- vLLM
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →