Home-Grown Heterogeneous Inference 'AI Token Factory' Solution Released by Qujing With Zhipu and Moore Threads

At the 2026 AI+ Conference, Qujing Technology, together with Zhipu and Moore Threads, released a "high-quality domestic heterogeneous-inference AI Token factory solution" that claims to close the model-inference-chip loop and reshape "token economics" through full-system heterogeneous coordination. Qujing founder and CEO Ai Zhiyuan said that as agents and long texts push call volumes toward exponential growth, the key to commercialization is attacking bottom-layer systems engineering to turn compute reserves into affordable, stable productivity. The solution uses Qujing's original heterogeneous collaborative inference: strong-compute chips handle the compute-dense Prefill stage and high-bandwidth-memory chips handle the Decode stage, with CPUs, GPUs, memory and high-speed networks all pulled into scheduling, achieving lossless precision and millisecond-level fault tolerance in production. Qujing said its domestic heterogeneous PD solution already runs in production carrying trillion-parameter models and high-concurrency traffic, and it is scaling a multi-source compute-supply network.

Source: Jintai Info / International Science and Technology Innovation Center, Nov 4, 2026 | 2026-11-04