SOLAR 10.7B punches above its weight class for a sub-11B model, showing solid instruction-following and reasoning capabilities that often rival larger models. It uses a depth-upscaling technique that merges layers from smaller base models, giving it an unusual architecture compared to standard transformers. The 4096-token context window is modest, and the non-commercial license limits production use cases.