Founders

The two co-founders of Yunjue

One serial operator with 20 years of management experience, one next-gen AI architecture researcher — together they set the "hardware × software × algorithms as one" product path Yunjue is on.
Peng Chao

Peng Chao · Co-founder & CEO

20 years of operating experience · hardware + software · domestic + overseas · full omnichannel

A career-long operator. Previously VP at Alibaba Group / CEO of Smart Connected (former GM of the Tmall Genie business unit); Head of E-commerce & User Operations at Honor; Country President of Huawei Consumer Business Group, India. Has led organizations of 1,000+ people and run full-stack P&Ls spanning hardware, software, e-commerce, and user operations.

  • Alibaba · CEO of Smart Connected (Tmall Genie) — led Alibaba's consumer AIoT business: scaled AliGenie to mass adoption, built commercial revenue from zero to scale, and improved the P&L from significant losses to near break-even.
  • Honor · Head of E-commerce & User Operations — drove Honor's online business to #1 share in the industry.
  • Huawei · Country President, India Consumer BG — full ownership of Huawei's consumer business in India: regulatory affairs, subsidiary management, India-localized R&D and supply chain, sales and marketing.
  • 14 years at Huawei — from branch sales to country president.
Qi Weizhen

Qi Weizhen · Co-founder & CTO

Next-gen AI architecture researcher · 3,000+ citations · h-index 17

A researcher focused on pretraining, agents, and inference acceleration. Three works defined cornerstones today's mainstream LLMs depend on: the pretraining paradigm (MTP), inference acceleration (KV Cache optimization), and the agent paradigm (LLM orchestrating multimodal tools).

  • ProphetNet (original MTP) — first author; the original multi-token-prediction pretraining paradigm, adopted as a core pretraining method by mainstream LLMs including the Deepseek series (V3, V4, etc.) and Qwen series (V3-Omni, V3.5, etc.).
  • EL-Attention (KV Cache optimization) — among the very first KV Cache optimization works in the field; the core idea aligns with later efficient-inference algorithms such as DeepSeek's MLA. Integrated into Huggingface Transformers; Microsoft Engineering applied the same technique to OpenAI's GPT-3 deployment, delivering a 5–10× output-speed boost.
  • Visual ChatGPT (core author) — 30K GitHub stars within a week of release; introduced the paradigm of an LLM orchestrating multimodal tools to complete complex tasks.
  • Engineering — solo-led full-pipeline pretraining of ProphetNet-X on 500 A100s and 5 TB of corpus; BANG (the field's first non-autoregressive pretrained model) powers Bing's <50 ms real-time response.
  • Academic roles — Area Chair, ACL / EMNLP ARR 2025; visiting research mentor, Zhongguancun Academy.
  • Education — BSc, University of Science and Technology of China; joint PhD program, Microsoft Research Asia.

Google Scholar →

See the techSee open roles