Large Language Models

Huawei is expected to unveil a significant AI inference technology at an upcoming forum, potentially reducing China’s reliance on High Bandwidth Memory (HBM). HBM is crucial for AI inference due to its high bandwidth and capacity, enabling faster access to large language model parameters. However, HBM supply constraints and export restrictions are pushing Chinese companies to seek alternatives. This innovation could improve the performance of Chinese AI models and strengthen the Chinese AI ecosystem.

The U.S. is launching “Project ATOM,” a strategic initiative to regain leadership in open-source AI amid growing competition from China, particularly Alibaba’s Qwen models. This U.S.-based non-profit AI lab will develop freely accessible AI models, supported by over 10,000 GPUs. Backed by industry leaders, the project addresses concerns about the U.S.’s lagging open-source contributions, highlighted by the dominance of Chinese-developed open-source LLMs. Project initiator Lambert emphasizes the need for coordination and funding, warning of potential U.S. decline in global AI influence if the initiative fails.

The 10th Scientific Data Conference highlighted Hygon DCU-powered innovations and collaborations with CAS, IHEP, and NAOC. Hygon and IHEP unveiled a scientific LLM solution leveraging Hygon’s DCUs and IHEP’s data. CAS showcased multimodal AI applications (“Zidong Taichu”). IHEP uses Hygon DCUs to manage big data in high energy physics, creating “Xi Wu,” a leading L2 model. Hygon’s DTK, DAS, and DAP optimize scientific software, achieving significant efficiency gains in astronomy and cryo-electron microscopy. Hygon aims to foster a Chinese technological innovation ecosystem.

The 2025 World Artificial Intelligence Conference (WAIC) showcased a shift in focus from Large Language Models (LLMs) to embodied intelligence, particularly humanoid robots. The event, touted as the “largest WAIC ever,” featured over 800 exhibitors. While humanoid robots drew huge crowds and high ticket prices, some insiders noted a “playing it safe” approach, prioritizing fundraising over groundbreaking innovation. LLM companies, while still present, appeared more subdued, focusing on practical applications and revenue generation, with many aiming for “¥1 billion in revenue.” Concerns linger about a potential bubble, especially in embodied intelligence.

Tencent’s Robotics X Lab and Futian Lab launched Tairos, a modular embodied AI open platform designed to advance robotics in China. Tairos uses large-scale AI models, offering a “plug-and-play” system with development tools and data services. The platform empowers robots to perceive, plan, and autonomously make decisions. Tairos includes multi-modal perception, planning, and perception-action joint models, improved by a cloud simulation environment and Tencent Games’ scene generation tech. Initial partners will help improve real-world applications of embodied AI tech through collaboration.

CAICT’s AI Institute launched security benchmark testing for code-generating LLMs, assessing risks and capabilities using a dataset of 15,000+ test cases across nine languages and various attack methods. The initial assessment of 15 Chinese models (3B-671B parameters) revealed varied security levels, with most exhibiting medium risk. Models showed weaknesses in scenarios involving malicious intent, highlighting vulnerabilities to cyberattacks. CAICT plans to expand testing to international models and develop mitigation tools, aiming to promote a secure LLM ecosystem.

GSR Ventures Managing Partner Zhu Xiaohu predicts that large language models (LLMs) will “devour” 90% of AI Agents. His comments, shared on Xiaohongshu, follow previous skepticism about embodied AI and highlight his firm’s bullish stance on the broader AI landscape, evidenced by investments in companies like Robopoet and LiblibAI. Zhu likened AI Agent startups to early internet webmasters, suggesting they learn from successful internet companies. His perspective sparks debate about the long-term viability of standalone AI Agents.

Huawei has open-sourced its Pangu 7B dense and Pangu-Pro MoE 72B large language models, along with Ascend-based inference technology. This move supports Huawei’s Ascend ecosystem strategy, aiming to accelerate AI research and application. The Pangu-Pro MoE 72B model shows strong performance, ranking highly on benchmarks for models under 100 billion parameters.