SHANGHAI, July 18 — At the 2026 World Artificial Intelligence Conference (WAIC), Chinese AI startup Moonshot AI unveiled Kimi K3, a massive open-source large language model featuring a 2.8-trillion parameter Mixture-of-Experts (MoE) architecture. The release positions it as the world’s largest open-weight model, marking a direct challenge to frontier closed-source systems from OpenAI, Anthropic, and Google.
The announcement came alongside a flurry of product releases from China’s largest technology companies, all converging on a single theme: artificial intelligence moving from benchmark-chasing showcases to practical, deployable tools for enterprise and consumer use.
Kimi K3: Scaling New Heights via MoE
Kimi K3’s reported 2.8 trillion total parameters (with a highly optimized active parameter count per token) would make it the first open-source model to cross the multi-trillion threshold, surpassing previous open-weight leaders including Meta’s Llama series and Alibaba’s Qwen models. In internal benchmarks, the model ranked closely behind Anthropic’s Claude 5 Opus and OpenAI’s GPT-5 / o3 — the most advanced closed-source systems publicly acknowledged — while outperforming competitors in programming, visual understanding, knowledge work, and long-context task handling.
Industry analysts noted that while parameter count alone does not determine model capability, the release signals Moonshot AI’s ambition to compete at the highest tier of the global AI landscape. The company, backed by major tech investors including Alibaba and Tencent, has rapidly expanded its research footprint since launching the original Kimi chatbot in 2023.
Tencent Unveils Embodied Intelligence Stack and ADP 4.0
Tencent used its WAIC platform to unveil what it described as a full-stack solution for embodied intelligence — AI systems that interact physically with the real world through robotics. The stack spans cloud infrastructure, model layers, development platforms, and application layers, aimed at accelerating the commercial deployment of autonomous robots.
Simultaneously, Tencent Cloud launched the overseas version of ADP 4.0, its enterprise-grade Agent Development Platform, alongside a “Top 10 Industries, 100 Scenarios” ecosystem program. The platform has already been deployed across more than 30 industries, covering intelligent customer service, knowledge management, and media production.
Lin Songtao, vice president at Tencent, characterized the shift as a move from “Q&A-style interaction to task-based collaboration,” noting that AI is evolving from being a personal assistant to becoming a core productivity engine for entire organizations.
Huawei Ascend 950 Ultra Node Debuts
Huawei presented the Ascend 950 Ultra Node for the first time at WAIC, describing it as the industry’s largest-scale ultra-node designed for data center construction, trillion-parameter MoE model training, and high-concurrency inference scenarios. The system features three core advantages: ultra-high bandwidth, ultra-low latency, and unified memory addressing.
The hardware reveal underscored China’s aggressive push to build domestic AI infrastructure and scale up cluster networking capabilities amid ongoing restrictions on access to advanced semiconductors and computing equipment from U.S. suppliers.
AI Governance Standards Take Center Stage
Beyond products, the conference placed significant emphasis on AI governance. Chinese regulators announced the initiation of a mandatory national standard for “Basic Safety Requirements for Agent Applications” — a framework intended to establish safety baselines for autonomous AI agent deployment across critical sectors.
Shi Yi, an academician at the Chinese Academy of Sciences, argued that embodied intelligence is forcing a fundamental reconstruction of technical standards. Robot perception, he noted, must evolve from visual-only systems to those incorporating tactile sensing, requiring entirely new frameworks for sensor calibration, platform interoperability, and real-world reliability.
The governance discussions reflected a broader narrative emerging at this year’s conference: that China’s AI industry is transitioning from a phase of rapid, unregulated model development to one focused on standardization, enterprise integration, and physical-world deployment.