DeepSeek

On November 2nd, the DeepSeek Harness team published a long paper on October 31, exploring Super-agent structures, noting that as models become more and more independent in their individual tasks, bottlenecks in multi-tasking collaboration are turning to attention. The team observed that one person could normally follow only 3 to 5 active sessions, and therefore wanted the system to take over the session management and focus on decision-making and acceptance. The team's programme is to divide duties into two levels: permanent Router, who coordinates and interacts with the outside world, and W..

On 31 October, the Information reported that DeepSeek had completed a second round of financing of approximately RMB 105,550 billion (approximately $15.8 billion) this month, from nearly 30 investors, valued at RMB 61.5 billion. It was reported that the era of telecommunication and Ningde was still the top two investors in DeepSeek. New investors in this round include the long-stamp storage and its Chairman, Chu Yiming, and Chinacore International. Vivo, American, American, well-informed, gleaned, travelling and little red books are also involved

On October 16, according to news reports today, DeepSeek recently consolidated the Harness team and the Web, App, Platform team into a “product department” responsible for all of the company's products, design and research and development, with Choi Wing as head of the product department. On 16 October, Tim Wing told the news that it was true. According to publicly available information, in March of this year, the team of Agent Harness from DeepSeek graduated from Zheung's computer department..

On 26 September, according to the news, OpenCode announced yesterday that the second phase of “Operation Cheepseek” would be launched, and the amount of $60 (note: the current exchange rate is approximately 403.5 yuan) originally provided by DeepSeek V4.1 Flash in OpenCode Go was adjusted to permanent validity. Before, OpenCode temporarily increased DeepSeek V4.1 Flash to four times the original standard, from $15 to 60..

On 24 September, DeepSeek recently released the internal Agent training and assessment sandbox infrastructure DSec. The paper was submitted to arXiv on 19 September, and Liang Wenbing was listed as the author. It describes the operational platform that has been put into production and not the release of new models. In the course of training, Agent needs to open a code warehouse, execute commands and call tools, and each mission must have an independent environment of sustainable preservation. DSec provides, through a unified SDK, a function call, a container, a mini-virtual machine and a complete virtual machine of four sands..

On 22 September, according to the Information report, the founder of DeepSeek, Liang Wenzine, stated at the investor closed meeting on 20 September that the use of a national chip, especially the Chinese chip, as a training model, was one of the company's most important current stakes and that the work “must succeed”. According to two sources, DeepSeek is expected to obtain a new shipment of Chinese training chips in the fourth quarter of this year or the first quarter of next year. According to the report, DeepSeek is finalizing a second round of financing with a target of 50 billion yuan yuan yuan, corresponding to the target..

On September 16th, the head of DeepSeek v4.1, Attention Chairman Liu Sheng and I had to bury my talent yesterday in a letter sent on September 14th, triggered a cyber-hot, received 100,000 + readings of the Twitter public number, went to the top of the hot list, and triggered much discussion on X. Liu Sheng and Yu wrote that AI had made rapid progress in the design of the algorithms he was working on, and that in just one year, AI had been able to help me check the files, read the code, find the bug..

On 9 September, according to the news, the head of the DeepSeek Harness team, Choi Tim Wing, publicly recruited approximately 150 people on 7 September. This number is intended mainly for engineers with 2 to 10 years of experience, who are also accepted as candidates and beginners; the place of work is mainly Beijing, with some of the posts available for Hangzhou. Recruitment is concentrated in service-end development and Agent flexible computing and development. The former covers large model research platforms, Agent framework components, research and development efficiency infrastructure, DeepSeek API, online services and data engineering; the latter includes..

On September 9, according to the news, DeepSeek will re-pricing the flash series at 12 noon on September 10, with the input price falling from $1.5 ton to $1 per million token, the output price falling from $4.5 to $4 per person per person per person per person per person per person per person per person per person per person per person per hour, and the cache price from 0.05 to 0.02 dollars per person per person per person per person per person per month. At the same time, the middle version of DeepSeek V4.1 Flash has been open-ended. API users participating in internal measurements need not change b..

On 4 September, according to information from sources quoted by AI Front, the former DeepSeek researcher, Wei, recently led a team to the Department for the Development of 100-degree Basic Models, where he was the head of the MMA. He joined 100 degrees in the first half of this year, before he was primarily responsible for research and development in multi-modular directions such as OCR. Wei is the first author of DeepSeek-OCR: Contexts Optifical Corporation and DeepSeek-OCR 2: Vision Capital Flow. Two..

Yesterday, 22 August, DeepSeek online experimental visual understanding model DeepSeek-V4-Flash-Vision-Exp was opened on the API platform. Users can call through the deepseek-v4-flash-vision-exp model name; officially, its pure text, Agent, reasoning and knowledge skills are essentially the same as the V4-Flash official version, while adding image understanding capabilities. In the official four multi-mode Agent assessment,..

On August 20, DeepSeek Harness published dsh-v0.1.0-rc.8. This update to the DeepSeek model adapter adds a configured raw picture request, and tasks like /goal, /plan can receive text input directly, @ menu can also cite files and existing sessions. This version also makes Claude Code and Codex sub-agents a needs-based Profile Bundle; in which Codex supports non-interactive access mode and multiple..

On August 18th, DeepSeek API introduced the Peak Valley Pricing Mechanism by formally adjusting the pricing strategy since August 17th at 00:00. The peak time period is set at 9-12:00 and 14:00-18:00 Beijing time, while the rest is free time and free time is half of the peak time. In the example of the flagship model DeepSeek-V4-Pro, the price per million token input per peak hour was raised from $3 to $9 and the output price from $6 to..

On August 1st, DeepSeek came on line yesterday with the V4-Flash official version of API, and the developer set the model name to be deepseek-v4-flash, which is the same as the original API call. This time only upgrades the API interface for V4-Flash without adjustments to the V4-Pro API, App and web-end models. The formal version of the model benchmark test exceeded the GLM-5.2 and the structure and parameter size followed V4-Flash-preview, with only re-training. I don't know

On July 23rd, the news that recently, the content of a four-hour investor ' s conference about DeepSeek founder Liang Wenfeng was circulated in the industry. Elsewhere recently released a compilation of the conference. During the meeting, Liang Wenbing exchanged views on the direction of DeepSeek, the AGI (General Artificial Intelligence) route, open source strategy, commercialization planning and competition in the AI industry. Below is a summary of the statements made by Liang Wenzeng during the nearly four hours of the exchange, organized by theme, to a total of 118 articles

On July 15, according to Bloomberg, the personal wealth valuation of the founder of DeepSeek, Leung Wenbong, has increased to $36 billion, up from $8 billion in Anthropic CEO Dario Amodei and $25.5 billion in Green Brockman, the co-founder and president of OpenAI, became the founder of the highest AI model in the country. Liang Wenzine's wealth growth is mainly due to his share of DeepSeek. DeepSeek was finished in June this year..

On 9 July, according to Krypton, China received an interview yesterday for the first group of "genius boys" and a former chief scientist, Pie AI, Li Boje, in response to DeepSeek's interview controversy and former investor Dudu's "roll run" allegations that had been the subject of extensive discussion. The interview started on July 6. In an X post, Lee Bojie said that in DeepSeek's two-side coding, he was repeatedly questioned by the interview officer over another screen that was set up on the conference interface because of his habit, and ended up on the other side saying that he could't prove himself innocent and terminate the interview..

On June 30, according to Blue Whale News, DeepSeek sent an e-mail to API users yesterday announcing that the V4 official version was scheduled to go online in mid-July and synchronized with the introduction of the Peak Valley pricing mechanism, at peak times two times the normal API price adjustment. Several developers have confirmed receipt of the mail. The peak time period is set at 9:00-1200 and 14:00-18:00, with the following pricing: V4 Pro peak: $0.05 / million tokens, cache undead input..

According to information received on 23 June, DeepSeek’s new Harness team is intensifying its recruitment efforts. The director of his Harness publicly announced that the DeepSeek Harness team was recruiting Harness Fellows, Harness Engineers and Harness Product Manager. In his job offer, Choi Tim Wing described the Harness team's objective as "Model + Harness = Agent": outside the model, around the context of the Agent product..

On June 17, the Information reported that DeepSeek had completed its first round of financing, raising more than 50 billion yuan yuan in total, updating the large model of national production and even the highest record of single-wheel financing in the entire Chinese AI industry to date. Of the publicly disclosed donors, Liang Wenfeng himself took the largest share of the purchase, contributing 20 billion yuan yuan; 10 billion yuan yuan yuan in telecommunication and investment; 5 billion yuan yuan in Ningdeh-era; and 3 billion yuan yuan each for Kyeongdong, Yeung and IDG capital. The rest..

News from June 5, yesterday, the founder and CEO Flo Criveello of the United States of America ' s AI Agent Creative Corporation wrote on X that the company had switched 100% Lindy traffic to DeepSeek V4 and discontinued the Claude model of Anthropic. Crivello claims that this switch could save companies millions of dollars a year and see improvements in performance in several core use scenarios. At the same time, he said he was still looking at the company of Anthropic..

On May 23rd, DeepSeek officially announced that the current DeepSeek-V4-Pro API discount of 2.5 would be converted to a permanent price reduction at 23:59 on May 31st. As of 1 June, the model API was adjusted to a quarter of the original price. DeepSeek API documents show that V4-Pro currently supports a 1M context, enter a cache outstanding price of $ 0.435 per million tokens and output price of $0.00..

On 23 May, according to Bloomberg, DeepSeek was engaged in a round of financing negotiations of about $10 billion (BND 70 billion), valued at about $45 billion. Two weeks ago, this financing came to the fore in the figure of RMB 50 billion (approximately $7.35 billion). According to the source, the Leung Wenfeng scheme, the founder of DeepSeek, may have contributed up to $20 billion, or 401 TP3T, of the total value of the round. If the deal is made, it'll be the largest start-up company in the history of China

On April 30th, DeepSeek launched the "Diagram Model" test yesterday, alongside the existing "fast mode" and "expert model" with full multimodular image understanding, not simple OCR text recognition. In real terms, DeepSeek is more accurate in general, and answers are available in half a second without opening the thinking mode. Common scenes such as film dramas, abstract pictures, and commodity maps are well identified and understood. It's even more important to think about the process: in addition to describing what's going on in the picture, there's going to be an active inquiry..