SenseTime Releases SenseNova U1.5-Lite with Unified Inference via OPD Distillation
SenseTime has released SenseNova U1.5-Lite, an open-source 8B parameter multimodal model that consolidates specialized expert models into a single unified model for inference. The model uses On-Policy Distillation (OPD) to merge capabilities in text rendering, aesthetics, and image editing without requiring a Mixture-of-Experts (MoE) router. This approach eliminates the computational overhead and complexity of Mixture-of-Experts (MoE) routing during inference while retaining the high performance of task-specific experts. It demonstrates a viable path for lightweight models to achieve commercial-grade capabilities in complex multimodal tasks like native 4K image generation and precise editing. SenseNova U1.5-Lite supports native 3-4k context lengths and features task-oriented reinforcement learning (RL) focused on instruction adherence, visual quality, and edit fidelity. It leverages JSON-structured supervision internally for precise control over spatial relationships and layouts, while maintaining a natural language interface for users.
## BACKGROUND
In traditional Mixture-of-Experts (MoE) architectures, a router dynamically directs inputs to specialized sub-networks (experts) during inference, which increases computational overhead. On-Policy Distillation (OPD) is a machine learning technique where a student model learns from a teacher model's outputs generated during its own active training rollouts, reducing distributional mismatch and allowing a single model to absorb multiple specialized capabilities.