BTC $64,384.71 -2.00%
ETH $1,798.47 -1.53%
BNB $588.38 -0.92%
XRP $0.9517 -0.71%
SOL $70.81 -0.48%
TRX $0.3025 -1.50%
DOGE $0.0584 -0.53%
ADA $0.1519 -0.80%
BCH $205.77 -0.75%
LINK $9.07 -1.43%
HYPE $66.36 +0.58%
AAVE $108.47 +2.48%
SUI $0.7161 -0.91%
XLM $0.1372 -1.76%
ZEC $885.93 -3.02%
AAPL $357.49 +0.08%
AMZN $264.37 +0.38%
GOOGL $294.77 +0.81%
MSFT $528.78 +1.42%
META $637.73 +0.27%
NVDA $160.18 -0.48%
TSLA $384.98 -0.59%
SNDK $838.68 -2.11%
INTC $57.03 -1.62%
SPCX $140.74 +2.07%
MU $743.84 -2.08%
AMD $511.91 -2.06%
BTC $64,384.71 -2.00%
ETH $1,798.47 -1.53%
BNB $588.38 -0.92%
XRP $0.9517 -0.71%
SOL $70.81 -0.48%
TRX $0.3025 -1.50%
DOGE $0.0584 -0.53%
ADA $0.1519 -0.80%
BCH $205.77 -0.75%
LINK $9.07 -1.43%
HYPE $66.36 +0.58%
AAVE $108.47 +2.48%
SUI $0.7161 -0.91%
XLM $0.1372 -1.76%
ZEC $885.93 -3.02%
AAPL $357.49 +0.08%
AMZN $264.37 +0.38%
GOOGL $294.77 +0.81%
MSFT $528.78 +1.42%
META $637.73 +0.27%
NVDA $160.18 -0.48%
TSLA $384.98 -0.59%
SNDK $838.68 -2.11%
INTC $57.03 -1.62%
SPCX $140.74 +2.07%
MU $743.84 -2.08%
AMD $511.91 -2.06%

deepseek

All
Article
Flash

Tencent Hunyuan released Hy4 flash, with total parameters of 224B and activation of 12B

Tencent officially released its first model named Hy4 flash, forming two tiers within the Hy4 series: flagship and low-cost Flash. The total parameters of Hy4 flash are 224 billion, of which 64 billion are Engram memory parameters, activating 12 billion during each inference, which is lower than Hy4's total parameters of 770 billion and 49 billion activation parameters, while still supporting multimodal and 1 million token context.Engram is a conditional memory technology previously proposed by DeepSeek, which stores some common text patterns in a quickly queryable memory table, allowing the model to retrieve information through table lookup, reducing redundant calculations. The input price for Hy4 flash is 0.5 yuan per million tokens, output is 2 yuan, and cache hit is 0.02 yuan, which represents a reduction of 12 times, 9 times, and 15 times respectively compared to Hy4's API list prices. Currently, Hy4 also has a limited-time discount of 50%, making the actual price difference less than the listed price difference.In Tencent's official testing, the collaborative editing score for PPT dropped from 78.9 for Hy4 to 74.0, with the cost per task decreasing from 1.714 yuan to 0.105 yuan, about one-sixteenth of the original. The score for automated tasks decreased from 56.2 to 52.2, with the cost per task dropping from 0.908 yuan to 0.038 yuan. In the Terminal-Bench 4.0 terminal operation test, Hy4 flash scored 19.2 points, while the flagship Hy4 scored 44.4 points, and the ProgramBench programming test dropped from 30.0 points to 16.0 points.

DeepSeek is hiring for operations and maintenance with a monthly salary of 30,000, while regular duty staff only earn five to six thousand

According to an interview with "Daily People," a data center operations engineer who has worked in Ulanqab, Inner Mongolia for 5 years stated that local AI data centers are still under large-scale construction, but the hiring demand at his company has significantly decreased. The company hired thousands of people in 2025, but by the time of the interview in 2026, only 3 new hires had been made.The engineer, who goes by the pseudonym Zhang Liwei, originally worked on an assembly line at an automobile factory. When he first switched careers, his internship monthly salary was 2,000 yuan, which increased to 4,500 yuan after graduation. After 5 years of job hopping and promotions, he currently earns over 10,000 yuan per month. In April of this year, DeepSeek was hiring senior operations engineers and senior delivery managers for its data center in Ulanqab, offering a monthly salary of 15,000 to 30,000 yuan, with 14 months' salary. These positions require extensive experience in data center operations and project management. According to Zhang Liwei's observations, local ordinary duty staff still earn around 5,000 to 6,000 yuan per month, while assistant engineers can earn over 10,000 yuan.The entry barriers are also rising. In 2024, some data centers were still hiring newcomers without relevant professional backgrounds; however, this year, some companies have required candidates to have a relevant degree and work experience, and there is even a 35-year age limit for basic duty positions. Zhang Liwei mentioned that a former colleague has started a training class for operations, charging nearly 20,000 yuan per student, and classes only start when there are 60 students, with a single session fee approaching 1.2 million yuan. The so-called internal recommendations are mostly just to help students secure an interview opportunity.

first_img Baidu launched the model trading platform Ciyuan, gathering 42 large models

According to "Du Jia," Baidu recently launched an open model service platform called "Word Element Institute" aimed at developers, enterprise clients, and model suppliers. It is positioned as a matchmaking trading platform connecting model suppliers and consumers, gathering 42 large models, covering text and video models such as DeepSeek, Kimi, and Zhipu GLM. Developers can connect to the Word Element Institute API once to switch between models from multiple vendors and be billed based on usage.The same model can be accessed by multiple suppliers. The platform displays pricing, initial token latency, throughput speed, 7-day success rate, and historical call volume, and supports both OpenAI and Anthropic protocols. Taking DeepSeek-v4.1-flash as an example, the platform connects to 6 suppliers, marking the price savings ratio on the page, while also displaying context window, launch time, deployment region, and supplier credit evaluation.The platform ranking is based on real call data and can be viewed for the last 24 hours, last 7 days, and last 30 days. Screenshots show a total of 398 million calls in the last 7 days, with a total token volume of 51.336 trillion, involving 68 ranked model resources, and an average initial token latency of 290 milliseconds. DeepSeek-v4.1-flash ranks first in both call volume and token usage. Participating suppliers include Bochuan Technology, Infinite Mind, Star Hub Word Element, and Zhipu Original Factory, with the platform showcasing resources and supplier information through credit evaluations and blockchain proof.

first_img Lisi Star Deep Fund's scale expanded to 7.52 billion yuan, continuing to acquire shares from DeepSeek

The Science and Technology Innovation Board Daily reporter learned from multiple sources that the venture capital firm Lisi Capital, which participated in the previous round of financing for DeepSeek, continues to secure an investment share in DeepSeek's new round of financing, with the participating entity being its subsidiary Tianjin Lisi Xing Shen Equity Investment Partnership (Limited Partnership) (Xing Shen Fund). As multiple listed companies disclose their funding progress and business information updates, the fund size has increased from the previously disclosed 2.64 billion yuan to 6.02 billion yuan, and was ultimately confirmed at 7.52 billion yuan, with the funds already paid in.Public announcements show that Zhihu announced on September 6 that it plans to invest 1.5 billion yuan through its wholly-owned subsidiary to participate in the Xing Shen Fund, which completed registration with the Asset Management Association of China on September 15; Tongchen Beijian disclosed in late September that it participated with its own funds of 50 million yuan when the fund size was 2.64 billion yuan. On October 9, Zhongwang Fabric announced its intention to invest 50 million yuan, at which time the subscribed amount had increased to 6.02 billion yuan. On October 21, Tongchen Beijian disclosed that it received notice that the fundraising scale had increased from 2.64 billion yuan to 7.52 billion yuan, the fundraising has been completed and all funds have been paid in; Perfect World, Te Bao Biotechnology, and others subsequently confirmed this scale. According to Qichacha, the Xing Shen Fund has a total of 41 partners, including the executive partner Tianjin Lisi Ming Tang and 40 limited partners, with a total subscribed capital of 7.52 billion yuan.The reporter compiled that 13 listed companies and their related funding entities have a total subscribed amount of 3.495 billion yuan, accounting for approximately 46.48% of the total fund size. Zhejiang Innovation Deep Intelligent Venture Capital Partnership has subscribed 1.4 billion yuan, making it the second-largest contributor; along with the two funds of Zhejiang Province, Ke Chuang Zhan Xin and Ke Chuang Wei Lai, each subscribing 50 million yuan, the three Zhejiang state-owned entities have a total subscription of 1.5 billion yuan, while Wuxi Chanfang Guosheng Innovation Development Equity Investment Partnership has subscribed 80 million yuan. A person familiar with the situation stated that Lisi is one of the market-oriented institutions that obtained a larger share, while other market-oriented institutions typically have quotas ranging from 3 billion yuan to 4 billion yuan. Zhihu previously announced that the Xing Shen Fund adopts a blind pool structure, and this person indicated that the actual allocation may still be highly concentrated in DeepSeek. The reporter had previously sought confirmation from Lisi Capital on whether the Xing Shen Fund participated in this round of financing, but had not received a clear response by the time of publication.

Qianwen had 158 million monthly active users in September, a year-on-year increase of 5077%, a decrease of about 9 million compared to June

According to the latest rankings from QuestMobile, in September 2026, Alibaba's AI application Qianwen App reached 158 million monthly active users, a year-on-year increase of 5077%, ranking first in growth among apps with over 100 million monthly active users. In the domestic AI native app rankings, Qianwen ranks second, surpassing DeepSeek's 123 million, but still below the first place, Doubao, which has 419 million.This growth is related to Alibaba's large-scale promotion. In November last year, the Tongyi App was renamed to Qianwen. This Spring Festival, Alibaba announced an investment of 3 billion yuan to launch a free order campaign, allowing users to order milk tea and takeout through AI. The campaign link was widely shared in WeChat groups, and the system experienced delays due to increased traffic. QuestMobile data shows that on the first day of the campaign, Qianwen's daily active users rose from about 7.06 million to 58.48 million.The year-on-year growth of 5077% was also influenced by last year's low base. Based on year-on-year data, Qianwen's monthly active users were about 3.05 million in September last year. By March this year, Qianwen's monthly active users had reached 166 million, rising to 167 million in June, and then falling back to 158 million in September, a decrease of about 9 million compared to June, with a decline rate of approximately 5.4%.

DeepSeek showcases the AI Workbench Harness Creator Mode

DeepSeek showcased the creator mode of the AI workstation Harness. Harness currently offers four modes: Standard, PTC, Minimal, and Creator. Among them, the Creator mode allows users to directly let AI develop plugins, modify software interfaces, and then install them for use in Harness without needing to write code themselves.Users only need to describe their requirements, and the AI will check Harness's development interface, write the plugin code, complete the installation, and check the running effects. The installed plugins can be saved for continued use the next time they open the application. In addition to adding new interfaces, it can also provide tools for the Agent, enabling it to perform operations it originally could not. For example, users can have AI create a video editing tool with a timeline, a paper reader that supports underlining and annotations, as well as generate poster design tools, interactive 3D models, or even change the Harness interface to a pixel game style.In the official demonstration, the AI created a floating tomato timer, writing over 600 lines of front-end code, completing the installation and operation, taking a total of 5 minutes and 24 seconds. DeepSeek also released the prompts for 7 demonstration cases. Some works were modified through multiple rounds of dialogue, and it is not possible to achieve the same effect by simply inputting a single sentence.
app_icon
ChainCatcher Building the Web3 world with innovations.