Quasar 438B prunes 117 experts per layer

Multiverse Computing’s Quasar 438B technical release keeps 148 of 265 experts per layer, a derived 117-expert or 44% reduction aimed at coding and agentic work. The current CompactifAI API serves the result in FP8, making Quasar a workload-specific compression bet rather than a drop-in generalist. Builders should benchmark their own traces before trading broader capability for a smaller routing footprint, as the five-task harness cost swing in coding-agent evaluations shows.