SOLAR 10.7B punches above its weight class for a sub-11B model, built using a depth-upscaling technique that merges layers from smaller pretrained models rather than training from scratch. It handles instruction-following and reasoning tasks with surprising coherence given its compact size, though its 4096-token context window keeps it better suited for focused, shorter interactions than long document work.