At the 2026 'AI Plus' Conference (2026 GAI+C), Qujing Technology, together with Zhipu and Moore Threads, released a 'high-quality domestic heterogeneous-inference AI Token Factory' solution. Leveraging system-wide heterogeneous coordination, it breaks the 'model–inference–chip' silos to reshape token economics — multiplying effective output per unit of hardware and sharply cutting overall cost.
The approach routes compute-intensive Prefill to strong-compute resources and memory-bandwidth-rich Decode to high-bandwidth resources, orchestrating CPU, GPU, memory and high-speed networking with millisecond-level lossless fault tolerance. It has already stably served trillion-parameter models under real high-concurrency traffic.
The move accelerates the closed loop from domestic chips and frontier models to scaled industrial deployment, turning computing reserves into affordable, stable and deterministic productivity for Chinese enterprises.
