From the source
Ant Ling released Ring-lite-2507, an upgraded version of its lightweight reasoning model Ring-lite (2506).
The model is built on a 16.8B Mixture-of-Experts architecture with 2.75B activated parameters and demonstrates improved performance across general text understanding, alignment, coding, logical, and agentic tasks.
The company also introduced a novel training approach called Constrained Contextual Computation Policy Optimization (C3PO) and released the training recipe and model weights publicly.
