November 4, 2026 AI Highlights: Wuzhen Summit Focuses on Open Source, Tencent Hunyuan Hy4 Released, DeepSeek V5 Planned on Ascend

The Wuzhen Summit is turning "open source and openness" into the central narrative of the AI industry. Models, compute infrastructure, AI agents, embodied AI, and governance frameworks are all advancing simultaneously, pushing AI from capability demos into scaled deployment.

AI Highlights — November 4, 2026

Today's AI news is dominated by the 2026 World Internet Conference Wuzhen Summit. Open-source ecosystems, large-model capabilities, domestic compute infrastructure, and AI Agent commercialization are the key themes.

Tencent Hunyuan Hy4 Released, Entering the Top Open-Source Model Tier

Tencent Hunyuan has released its next-generation large language model Hy4, significantly expanding data and training scale during pre-training and post-training. Hy4 introduces visual understanding for the first time and improves instruction following, long-document comprehension, and long-horizon agent capabilities.

Enflame Technology has completed Hy4 adaptation, using operator fusion, memory optimization, intelligent scheduling, and multi-level KV Cache to support larger models and million-token ultra-long-context inference.

DeepSeek V5 Planned for Ascend-Based Pre-Training

Industry sources say DeepSeek's next-generation model plans to run pre-training on Huawei Ascend chips. Ascend has previously been used more for inference; moving pre-training onto domestic hardware would mark a more substantial step for "domestic chips training domestic models."

This matters for global AI infrastructure developers: domestic compute stacks are moving from inference-ready to training-capable, which could reshape how model APIs, inference services, and multi-region deployments are adapted in the future.

Wuzhen Summit: China's Generative AI Users Surpass 600 Million

According to the China Internet Development Report 2026, China had 602 million generative AI users by the end of 2025, with a penetration rate of 42.8%. Daily token calls reached 140 trillion in March 2026, growing more than 1,000 times in two years.

China's enterprise AI Agent market reached RMB 21.2 billion in 2025, while the embodied AI market reached about RMB 915 billion, up 20.4% year over year. AI is shifting from "answering questions" to "executing tasks."

Agent Commercialization: From Demos to Delivery

Jazzyear's 2026 Agent Commercialization and Competition Research Report argues that Agent competition is expanding from single-model capability to end-to-end task delivery. Models determine whether a task can be completed; business context, permissions, tool connections, and exception handling determine whether it can be delivered reliably.

Citing BCG data, the report estimates Agentic AI's share of AI value will rise from 17% in 2025 to 22% in 2026, and reach 39% by 2030.

Other Notable Updates

  • Huawei Ascend 950 SuperPoD received the World Internet Conference Leading Technology Award, integrating 1,024 accelerator cards per node for trillion-parameter training and high-concurrency inference.
  • SenseTime released SenseNova 6.8 Flash, supporting cross-task and cross-session long-term memory, with generation extended to dynamic PPTs, webpages, SVG, Three.js, and Blender 3D.
  • Tencent Hunyuan launched Hyra 2.0, a research agent that moves general-purpose agents toward scientific model development, research workflows, and software tool creation.
  • NetEase Qiyehui upgraded its enterprise AI Agent platform "King Crab", debuting a smart badge and introducing the "Wandian" knowledge engine and Agent Team capabilities.
  • Apple opened the Siri AI preview to all developers, while Apple Intelligence integrates Alibaba's Qwen model.
  • MIT proposed "weight-space programming", improving a small model's math capability by about 6x by modifying roughly 0.016% of its weights.
  • CrowdStrike acquired Fiddler AI and DreadNode for million and introduced an AI Agent "kill switch" for observability and runtime protection.

Today's Takeaway

The Wuzhen Summit confirms that AI competition has moved beyond single-model benchmarks into a broader contest over open-source ecosystems, compute infrastructure, Agent delivery, and governance. For global developers, open-source models lower entry barriers, but product success will increasingly depend on task orchestration, permission control, tool integration, cost management, and multi-region deployment.

Last updated November 4, 2026

Where to go next