DeepSeek · Full chronicle and history

On October 28, The Information reported that the latest models from DeepSeek and Moonshot AI are suspected of using output from OpenAI's GPT-6 Astra for distillation training. In September, OpenAI traced distillation activities that began in July to individuals associated with Moonshot AI. Anthropic stated that it has reduced the data acquisition for distillation attacks against Claude by approximately 90% to 99%.

On October 27, Weibo users reported that when they asked what "inappropriate remarks" Mbappé had made, multiple Chinese AI apps, including Doubao, DeepSeek, and Xiaowei AI, all refused to answer.

Researchers released a preprint paper testing major language models for gender bias, finding huge differences among models: when asked whether it is acceptable to abuse women to prevent a nuclear apocalypse, DeepSeek strongly agreed, while Claude, GPT, and Llama all strongly disagreed.

On September 30, Valerio Capraro, a researcher at the University of Milano-Bicocca in Italy, released test results: when presented with the scenario of "whether one person can be abused or tortured to prevent a nuclear apocalypse," DeepSeek V4-Flash answered "strongly agree" in each of 50 repeated queries, while under the same conditions both Llama 4 Scout and Microsoft Copilot chose "strongly disagree."

According to The Wall Street Journal, ZRON built an analytics dashboard that integrates public information and private data, introducing large models such as Alibaba and DeepSeek, which can automatically generate intelligence reports and organize event timelines, improving the readability and practicality of intelligence products.

On September 11, 2026, the US AI giant Anthropic released a long-form investigative report detailing how Chinese AI companies such as Moonshot AI and DeepSeek, while distilling own models from US models like Claude, allegedly sent real internal prompts and sensitive data of users from the People's Liberation Army, police, and state-owned enterprises to US AI platforms, sparking strong concern over model distillation security and data leakage.

On September 10, Anthropic released a threat intelligence report stating that Chinese AI companies, including DeepSeek and Moonshot, used third-party API proxy services to bypass regional restrictions and call Claude. The report also noted that these relay services may obtain model access through methods such as fake accounts and stolen API keys.

Taiwanese cybersecurity research firm TeamT5 released a report stating that after Chinese state-linked hacker groups integrated open-source AI models such as DeepSeek into their attack workflows, the number of attacks more than doubled; DeepSeek has become hackers' tool of choice due to its strong performance, low guardrails, and low cost.

In August 2026, DeepSeek, regarded as a representative Chinese artificial intelligence company, announced a substantial increase in the pricing of its large model API calls.

In August 2026, remarks allegedly made by DeepSeek founder Liang Wenfeng at a closed-door meeting circulated online, in which he said that the core bottleneck currently limiting corporate development lies in computing power cards and chip supply.

The day after Fang Xinghai was reported, Liang Wenfeng notified potential investors for the second round of financing to pause, with the valuation for the relevant round approaching $70 billion.

Chinese officials have reportedly told companies such as Alibaba, ByteDance and DeepSeek that they may soon be permitted to purchase some H200 chips, with the quantities still being determined.

Reuters reported on June 17, 2026 that, out of consideration for avoiding escalating U.S.-China relations, the U.S. government has delayed formally adding more than 100 Chinese entities — including AI company DeepSeek and chipmaker ChangXin Memory Technologies — that had already received interagency approval to the export control Entity List.

Bloomberg reported on May 26 that some of the top AI talent at private companies such as Alibaba and DeepSeek must obtain official approval before leaving the country, with the list selected based on an individual's strategic importance to the country rather than solely on rank or employer.

According to a Bloomberg report cited by the program, government departments require senior AI researchers and executives at companies such as Alibaba and DeepSeek to obtain prior approval before leaving the country.

The claim that DeepSeek trained a top-tier model for $5.6 million, along with its technical report, has upended Nvidia's $589.1 billion market capitalization.

On January 20, 2025, Chinese AI startup DeepSeek officially released the open-source inference model DeepSeek-R1, shaking the global tech community and sparking discussion of chip export controls due to its low cost and high performance.

DeepSeek suddenly became popular in 2025, and the program claimed it was used to prove that China's artificial intelligence industry had achieved success during Li Qiang's tenure.

In early 2025, after Chinese AI DeepSeek became popular overseas, tests by multiple research institutions found that it systematically filters sensitive topics such as Tiananmen and Xinjiang Uyghur human rights, and that strict ideological alignment remains even when running offline.