The Compute Brief · Aug 6, 2026

China’s model and chip stack sees fresh activity

ByteDance, Huawei, Alibaba and DeepSeek all made moves around models, chips and deployment, while NVIDIA reportedly deepened talks with Chinese base-station makers on edge AI systems.

• ByteDance is discussing training a model with more than 5 trillion parameters, and CEO Liang Rubo said the company will stick with self-developed LLMs even if it lags in the short term.

• Huawei said Ascend now has 0-day support for Ant’s Ling-3.0-flash model and introduced the CANN PyPTO operator programming framework for faster deployment and tuning.

• Huawei also said Ascend supports reinforcement-learning training-inference consistency, with tests on Qwen3-30B MoE showing 20% to 60% performance gains in Eager mode.

• Alibaba DAMO Academy opened 15 research projects spanning AI chips, a new CPU architecture, medical multimodal agents and AGI decision-making, and launched top-talent recruitment for 2027 graduates.

• DeepSeek said it plans a broad and significant increase in API prices.

• NVIDIA is reportedly in talks with Chinese base-station makers, including Jiaxian Communication, to co-develop 6G stations for overseas markets and target trial networks in 2027 or 2028.

• In Hong Kong trading, large-model stocks rose even as the broader tech and chip sector weakened, with MINIMAX-W up more than 14% and Zhipu up more than 4%.