deepseek

According to "Du Jia," Tencent has launched the internal testing of "Tencent Cloud Drive," with some features still being iterated, and plans to release an App version. This product focuses on file storage, data management, and AI applications, with a core positioning as a unified memory base for multiple Agents, aiming to connect various AI applications and office tools, unify and carry context, allowing multiple Agents to collaborate based on the same information.Practical tests show that the internal test version retains basic entry points such as the homepage, files, recycle bin, and storage space, with a floating AI assistant panel on the right side. The core highlight of the product is its "connection" capability, supporting binding with Tencent Docs, WorkBuddy, and integrating third-party Agents like DeepSeek Harness and OpenClaw. After authorization, the Cloud Drive AI assistant can read application files, automatically sync updates, and directly call information as a knowledge base for retrieval, summarization, and content generation. The binding page will clearly inform users of the scope of permissions and the rules for AI credit consumption.The report points out that, unlike traditional Cloud Drive AI's document summarization, this system fully connects documents, Cloud Drive files, and Agent workspaces, shifting the Cloud Drive from a file repository to a data memory platform in the AI era.

The semiconductor research organization SemiAnalysis has released the actual test results of Huawei's Ascend 950PR. The organization selected 1,055 sets of commonly used matrix calculations for large models. Huawei's own underlying program has allowed the 950PR to achieve 86% of its theoretical computing power, higher than NVIDIA's H100 at 75.5%; however, the actual computation speed is still only 43% of the H100, and only 21% compared to the relatively newer B300. This set of results has not yet utilized the newly ported operators from DeepSeek.DeepSeek is working on reducing data movement and waiting from the software side. In attention computation, after adjusting storage with FlashMLA, the two sets of data that originally needed to be read separately can now be retrieved at once; TileLang allows the chip to compute while preloading the next batch of data. SemiAnalysis measured that automatic scheduling has reduced the time for some small computing tasks by up to about 40%. DeepSeek has also ported DeepEP, which is responsible for MoE expert communication, but communication and computation will still compete for memory bandwidth, and Huawei plans to continue improving the firmware.Last year, DeepSeek attempted to train R2 using Huawei's Ascend chips, but even with Huawei engineers on-site for assistance, it could not be completed due to issues related to chip stability, inter-card communication, and software support. This year, both parties continue to collaborate, and by the end of September, DeepSeek focused on publicly releasing six sets of Ascend adaptation code. In October, it was reported that Huawei is adjusting the next generation of chip designs according to DeepSeek's model requirements. The 950PR tested this time is mainly aimed at inference input processing, while DeepSeek is preparing to use the 950DT for large-scale pre-training as early as the end of the year. Previously, both parties had conducted tests at the thousand-card level, but to enable tens of thousands of chips to complete a round of large model training continuously, more complex coordination and stability issues still need to be resolved.

The Hurun Research Institute released the "2026 Hurun Rich List." This year, a total of 1,529 entrepreneurs with personal wealth exceeding 5 billion yuan made the list, an increase of 95 from last year, setting a new historical high. The total wealth of the listed entrepreneurs is 32 trillion yuan, a year-on-year increase of 7% (2.1 trillion yuan).Zhang Yiming, the founder of ByteDance, became China's richest person for the second time with 670 billion yuan, breaking the record for the wealth of China's richest person. Zhong Shanshan, the founder of Nongfu Spring, saw his wealth shrink by 120 billion yuan, falling to second place with 410 billion yuan. Liang Wenfeng, the founder of DeepSeek, made the list for the first time, entering the top ten with 280 billion yuan.

According to the report by the Science and Technology Innovation Board Daily, citing the Daily Economic News, the latest data from OpenRouter shows that last week (from October 19 to 25), the total call volume of global AI large models was 188 trillion tokens, a month-on-month increase of 14.63%. The top five global call volumes last week were all Chinese AI large models, with DeepSeek V4.1-Flash-1017 ranking first, having a weekly call volume of 33.6 trillion tokens, a month-on-month increase of 801%.Among the models listed during the same period, the weekly call volume of Chinese AI large models was 93.54 trillion tokens, a month-on-month increase of 18.14%; the weekly call volume of American AI large models was 22.47 trillion tokens, a month-on-month increase of 29.44%. The weekly call volume of Chinese large models has exceeded that of the United States for twenty-six consecutive weeks, ranking first in the world.

On October 23, B.AI announced the domestic model brand Token weekly ranking from October 16 to 22: DeepSeek continues to lead, with Xiaomi MiMo closely following, with only a difference of 16.54 B Tokens. Xiaomi has seen growth for 4 consecutive weeks, and the platform Token usage ratio has increased by 8.13 percentage points.The competition in the mid-tier is heating up, with MiniMax experiencing a month-on-month increase of over 5,200%, becoming this week's growth dark horse. Qwen saw a month-on-month growth of 80.86%, narrowing the gap with the third-place Zai to 0.18%. Popular model discounts are ongoing: DeepSeek-V4.1-Flash-1017 enjoys a 40% discount, Qwen3.8-Flash enjoys a 20% discount, and GLM-5.3-Flash enjoys a 30% discount. B.AI continues to gather high-quality models, helping developers unlock diverse AI capabilities at low costs. Welcome to log in to b.ai to explore.

According to People's Finance News, regarding DeepSeek's comprehensive open-sourcing of its internally used AI inference framework DeepInfer, Zhou Hongyi, the founder of 360, stated that this reflects a kind of technological confidence and, more importantly, a philosophy of open-source inclusiveness.Zhou Hongyi also mentioned that looking deeper, this indicates that the struggle for technological sovereignty in the AI era has arrived. He revealed that 360 has developed a sovereign AI operating system, which is currently in use in some government and enterprise units.

Early investment institution ZhenFund has launched Ask AI on its official website. Entrepreneurs can directly inquire about which teams ZhenFund has invested in, what fields they are currently focusing on, and what activities are suitable for them.ZhenFund has compiled publicly available information about its investment team, invested companies, areas of interest, entrepreneurial activities, and contact information into a bilingual zhenfund.md. After users click on any AI entry point, the website will redirect to Kimi, ChatGPT, Claude, or DeepSeek, and pre-fill the prompts, allowing the model to read the document before answering questions.

CCTV's "Yuyuantan Tian" disclosed that after China launched the AI technology export registration management on October 16, DeepSeek and MiniMax have successfully completed their registration and the related technology has been open-sourced and authorized for overseas use.Currently, the new catalog is still open for comments, and a temporary registration system is being implemented. The Ministry of Commerce previously clarified that technologies that have entered the public domain through open-source are not included in the management scope, and there are automatic licensing and other facilitative measures for new technology open-sourcing, which can basically be reported and released immediately. Providing model services overseas through API, web pages, and apps is also not within the scope of this large model technology export management.

According to the "Science and Technology Innovation Board Daily," the artificial intelligence company DeepSeek has introduced multiple industrial and financial investors in its new round of financing, including Legend Holdings' investment platform Junlian Capital, semiconductor-focused Stone Creek Capital, and Geely, among others.At the same time, DeepSeek has hired CITIC Securities to provide advisory services for its IPO and has introduced a new Chief Financial Officer in September this year. Junlian Capital and Lenovo Group are both part of the Legend Holdings system, with Lenovo Group being the core industrial company under Legend Holdings, and Junlian Capital serving as its important investment platform. Junlian Capital's participation in DeepSeek's financing also signifies that capital from the Lenovo system has officially entered DeepSeek's shareholder camp.

According to Bloomberg, DeepSeek raised approximately 100 billion yuan (about 15 billion USD) in its latest round of financing, making it one of the largest financings in China. Insiders said that the company, headquartered in Hangzhou, raised its financing target at least twice during negotiations with investors, first increasing it from 50 billion to 80 billion, and finally reaching the aforementioned amount.CATL and Tencent are among the largest investors in this round. This financing will be used to support DeepSeek's competition globally with American AI companies such as Anthropic and OpenAI.

DeepSeek announced the full open-source of its internal inference framework DeepInfer, and simultaneously open-sourced the production-level token encoding and decoding component deepseek-recipe. DeepInfer integrates high-performance operators and a deeply optimized inference engine, supporting distributed inference, speculative decoding, SWA Bounded Replay, and structured output, compatible with NVIDIA and Huawei Ascend platforms.DeepSeek stated that it has implemented high throughput and low latency large-scale inference services for DeepSeek V4.1 Flash based on DeepInfer, and future inference code and optimizations for new models will continue to be updated here. deepseek-recipe is fully aligned with the processing logic of the DeepSeek official API, aiming to ensure that the community receives an experience consistent with the official deployment. The open-source repository and technical report have been released simultaneously.

According to Beating reports, Tencent announced the opening of Yuanbao Search to developers through Tencent Cloud, supporting API, MCP, and Skill access. Tencent disclosed that this search service has an index of hundreds of billions, with an average daily call volume of hundreds of millions, and has supported over 50 internal businesses including WorkBuddy, ima, and Tencent Video. The underlying technology of Yuanbao Search comes from Sogou Search, integrates Tencent's content ecosystem, and connects to 157 authoritative sources across 13 industries including finance, healthcare, and law.Records from the Tencent Cloud Internet Search API show that this service went live on July 1, 2025. Alibaba Cloud launched its independent Internet Search API on March 20 of the same year; Volcano Engine upgraded its existing service to Doubao Search in June this year and officially released it to the public in July; DeepSeek opened its self-developed search engine's Search and Fetch API on October 17. The prices for Yuanbao Search's Express, Standard, and Premium versions are 14 yuan, 20 yuan, and 26 yuan per thousand calls, respectively, with an additional monthly free quota of 2,000 calls; the Express version only supports public web sources. The price for DeepSeek Search API is 5 yuan per thousand successful requests, Alibaba Cloud OpenSearch Lite is 12 yuan, and the Pro version was reduced to 18 yuan on October 15.

The AI model evaluation platform Arena announced that DeepSeek-V4.1-Flash-20261017 (Max) has ranked first among open-source models in Agent Arena and 11th overall. Based on approximately 7,500 real usage sessions, this model shows a comprehensive net improvement of about 5.1% compared to the average model, with a median single-task cost of about $0.20.This version ranks first among open-source models in four metrics: task success confirmation, user positive and negative feedback, reliability of Bash command execution, and guideability, with the task success confirmation metric ranking 8th overall.

The spokesperson of the Ministry of Foreign Affairs, Mao Ning, hosted a regular press conference. A reporter from AFP asked whether the U.S. government has requested federal agencies to stop using Chinese artificial intelligence products like DeepSeek, citing concerns about the potential transfer of sensitive data, and inquired about China's comments.Mao Ning stated that China has always opposed the generalization of the concept of national security and the politicization of technological and economic issues, and will firmly safeguard the legitimate rights and interests of Chinese enterprises. Previously, the U.S. government prohibited civilian agencies from using DeepSeek products.

According to The Information, DeepSeek is collaborating closely with Huawei to jointly design the next generation of AI chips, aiming to build models that do not rely on NVIDIA's computing power. Insiders revealed that Huawei engineers have entered DeepSeek's office to understand its model architecture and algorithm requirements, and to test how the chips handle data transmission and compression, with the first compatible chip targeting Ascend 950DT.The report pointed out that this move comes as China pushes to reduce its dependence on American chips. Limited by TSMC's foundry services and U.S. export controls, Huawei's Ascend 950DT is planned for release in 2027, but mass production may be delayed until 2028. To address the chip shortage, DeepSeek is using the TileLang programming language to bypass CUDA restrictions, while still attempting to acquire the banned NVIDIA Blackwell chips through third-party data centers. In addition, DeepSeek is forming a chip team and has begun testing clusters with over 1,000 processors.

The U.S. Department of Homeland Security (DHS) instructed U.S. civil agencies on Monday to remove products from the Chinese AI company DeepSeek from government IT systems, as well as any third-party services built on that model, and to prohibit new procurements. The notice stated that using DeepSeek could lead to the transfer of sensitive data to China and could generate intelligence targeting U.S. individuals and institutions. DHS indicated that these models constitute "multiple threat vectors."The ban targets not only DeepSeek's own applications but also covers third-party services based on its development. A DHS memorandum from July was made public on Monday, which had previously requested agencies to monitor the risks related to DeepSeek in the traffic in and out of government websites. The U.S. Department of Commerce, the Navy, and the states of New York, Tennessee, and Virginia had previously restricted DeepSeek. The Federal Communications Commission (FCC) will vote on October 29 to decide whether to ban all Chinese laboratories from testing electronic products used in the U.S.DeepSeek is a startup based in Hangzhou, and its open-source model attracted widespread attention on January 27, 2025, with development costs of less than $5 million, leading to a 16% drop in Nvidia's stock price that day, evaporating about $600 billion in market value. In April 2025, the U.S. House of Representatives' Special Committee on China Issues stated that DeepSeek might secretly send U.S. user data to the Chinese government and may have used about 60,000 Nvidia chips, of which 20,000 were subject to U.S. export controls. The V4-Pro model preview released by DeepSeek in April this year was priced about 98% lower than OpenAI's GPT-5.5 Pro and was made available for free download and modification under the MIT license; this week it also released an updated version of the V4.1-Flash model focused on efficiency optimization.

According to Beating reports, Lei Feng Network interviewed several headhunters and found that major domestic internet companies are tightening their recruitment quotas but are still willing to break salary limits for top AI talents. Headhunters revealed that this year, some outstanding fresh graduates in foundational model directions have annual packages reaching 2 million yuan; ByteDance Seed and Xiaohongshu's highest offers have reached 4 million to 5 million yuan, while Baidu is in the range of 2 million to 3 million yuan.The high salaries for talent are exacerbating internal salary discrepancies. An internet professional stated that some employees in foundational models have lower job levels but earn more than their higher-level colleagues. Similar situations emerged as early as 2017; an experienced engineer recalled that at that time, a returnee fresh graduate who joined a year later and was at a lower job level had a monthly salary 10,000 yuan higher than his own, leading to a monthly turnover of about 10% in a team of over thirty people.Headhunters noted that application-side positions are shrinking, while foundational model teams still lack individuals who can truly train large models. Coding Agents, model evaluation, pre-training data, and Agent infrastructure have become new focal points for talent acquisition; DeepSeek has recently been expanding its recruitment for backend developers and Agent Infra engineers, while the hiring enthusiasm in areas like AI search has decreased. Companies are not only looking at academic qualifications and papers but also place greater importance on whether candidates have actually trained large-scale models in leading teams; the salaries and responsibilities of scarce talents can be negotiated separately, while competition among ordinary AI job seekers is more intense.