Moore Threads Unveils Three 'AI Factories' Framework at WAIC 2026, Completes Full MoE-236B Training on Domestic Chips

At the 2026 World Artificial Intelligence Conference (WAIC) in Shanghai, Chinese GPU maker Moore Threads emerged as a key player in the AI infrastructure track. Founder, Chairman, and CEO Zhang Jianzhong unveiled a three-'AI Factory' framework built on the company's KUAE intelligent computing cluster, targeting the full chain of computing needs from model training to inference deployment and agent deployment.

Zhang noted that frontier foundation models now iterate every two months on average, with token consumption growing rapidly and computing demand rising exponentially. The three AI factories — the Model Training Factory, Token Production Factory, and Agent Production Factory — are designed to support diverse computing scenarios spanning training to inference and digital to physical worlds.

On the training front, the Model Training Factory achieved several breakthroughs. The KUAE cluster maintains up to 95% linear scaling efficiency, with training precision on par with international mainstream GPUs and highly consistent loss curves. The cluster supports checkpoint-based training resumption, with effective training time exceeding 90%. Moore Threads has completed multiple 'national chip training national model' projects, including training a MoE-236B foundation model from scratch, validating domestic chips' capability for ultra-large-scale, long-duration stable training.

For inference, the Token Production Factory deeply integrates with mainstream frameworks like SGLang and vLLM, covering LLMs, vision, speech, DiT, and video generation scenarios. Moore Threads achieved 'release-time compatibility' with major models including DeepSeek, MiniMax, GLM, Kimi, and Qwen. The company also introduced a PD-separated heterogeneous inference solution based on the MTT S5000, splitting high-compute Prefill pools from high-bandwidth Decode generation pools to significantly improve cost-performance ratios.

In the agent deployment segment, Moore Threads demonstrated its self-developed digital agent 'Xiaomai,' equipped with over 60 skills and supporting cross-app control across 38 applications. The company also showcased the MT Lambda full-stack embodied intelligence simulation platform, building a complete solution from underlying compute to core engine and upper-layer frameworks for producing physical-world agents.

To support Xiaomai's evolution, Moore Threads launched two AI-native devices — the MTT AIBOOK computing notebook and the MTT AICUBE home AI hub — creating a cloud-edge-end collaborative intelligent computing matrix. The entire ecosystem is built on the open-source MUSA architecture.

Moore Threads also disclosed several partnership milestones. Jijia Vision completed training of driving world models and physical AI models based on the MTT S5000. Peking University's EvoPhys project achieved full-stack native training of the 5D world model EvoPhys-World on the MTT S5000, topping the WorldScore 'world generation' ranking for 37 consecutive days. JD Cloud announced continued collaboration with Moore Threads on joint JoyAI model development and enterprise-level inference platforms to unleash the industrial productivity of domestic computing power.