Qwen3 (232Bx22B 'T') is a hybrid instruct + reasoning text model based on a sparse mixture-of-experts architecture. It balances efficiency, performance and quality in an inference service configured for high throughput.