Ox-Alpha Unmasked as GLM-5.3-Flash

Z.ai just confirmed the anonymously tested "ox-alpha" model is GLM-5.3-Flash: 320B params MoE, only 18B active, and it's already closing the gap with Claude Opus 4.8 on coding and agentic benchmarks. Efficient MoE architectures are increasingly matching flagship dense models at a fraction of active compute.