Open cooperation built around open source has become the main thread of artificial intelligence’s global expansion, according to an annual list of industry trends from the China Academy of Information and Communications Technology (CAICT), a research institute under the Ministry of Industry and Information Technology.
CAICT and the Artificial Intelligence Industry Alliance (AIIA) released the “2026 AI industry top ten keywords” on 15 October at the 2026 Artificial Intelligence Industry Conference in Jinan, where CAICT vice-president Wei Liang presented them. CAICT sums up this year’s theme as “accelerated diffusion”: whether AI can spread to more sectors at a cost people can afford and in a way they can trust. The ten are model self-evolution, agent engineering, agent devices, real-world training for embodied AI, chip-model coordination, high-value use cases, AI-driven automated labs, high-quality token services, frontier-model safety risks and “open and inclusive” (kaifang puhui).
Under the last of these, CAICT says AI open source has moved from releasing models to opening up the full stack. It says six of the ten most-downloaded models on Hugging Face in September were Chinese, and that Chinese products made up ten of the world’s top 20 AI applications by users in July. It does not say how either ranking was compiled. A SouthwardTech check on 19 October found seven Alibaba Qwen models among the ten most downloaded in Hugging Face’s text-generation category, which ranks models by recent downloads.
CAICT says AI companies worldwide are moving in the same direction: open source in exchange for ecosystems, “Day 0” globalisation for room to grow, localised products for trust and early compliance for market access. With more than 80 countries having published AI strategies and more than 40 new AI laws passed, it says, firms with global ambitions need to weigh, from the start of product design, how they allocate data, technology, talent and capital at home and abroad. Separately, on 16 October China’s commerce and science ministries issued draft export controls on AI technology.
On tokens, CAICT puts China’s average daily consumption at about 500 trillion in June, nearly four times the December 2025 level. The National Data Administration had given about 100 trillion for the end of 2025 and more than 140 trillion for March 2026; CAICT does not say how its figure was measured. It says grey-area practices are increasing, including quietly swapping in cheaper or cut-down models, over-reporting token counts and arbitrage through pooled accounts.
On domestic computing, CAICT cites China Telecom’s 10,000-chip cluster in the Greater Bay Area, built on Huawei Ascend supernodes, and Sugon’s 100,000-chip Sugon 8000, which it says is now in service. Under frontier-model safety risks, it says that in its own tests the rate at which an agent running on DeepSeek-V4-Flash triggered harmful behaviour rose by 12.5 per cent after one round of tool-skill self-evolution. It does not give the baseline or say whether the rise is in percentage points. It calls for international safety cooperation through channels such as the China–US AI dialogue, which the two governments agreed to set up in May 2026.
Related