deepseek

DeepSeek announced the new pricing for the V4 API, effective August 17. The V4-Pro peak period price for cache misses on input has increased from 3 yuan to 9 yuan per million tokens, and the output has risen from 6 yuan to 27 yuan, representing increases of 3 times and 4.5 times, respectively.During off-peak hours, although the price is only half of the peak rate, it is still more expensive than the current rates. The V4-Pro input is 4.5 yuan, and the output is 13.5 yuan; the V4-Flash input is 1.5 yuan, and the output is 4.5 yuan.Peak hours are from 9:00 to 12:00 and 14:00 to 18:00 Beijing time. The remaining hours are charged at off-peak prices.

B.AI platform officially announces that the DeepSeek V4 series Web Chat and API have been fully upgraded to the latest official version, covering two major versions: V4-Flash-0731 and V4-Pro-0813. Among them, V4-Flash-0731 has comprehensively enhanced Agent collaboration, code generation, and multi-tool invocation capabilities through specialized post-training while maintaining the original efficient architecture and model scale, resulting in significant improvements in execution efficiency and task completion; V4-Pro-0813 has shown significant improvements in test scores such as DeepSWE and CyberGym, further consolidating its leading advantage in complex engineering and agent tasks.Both models retain the original model ID and calling interface, allowing developers to switch seamlessly without any configuration changes, maintaining consistent compatibility and stability. Starting today, the new capabilities are fully open for experience, and you are welcome to log in to the B.AI platform to use them immediately.

According to blogger Max For AI, DeepSeek Harness (DSH) will officially start public testing on August 13. The final internal test version will be pushed tonight for developers to complete the final plugin compatibility work, and the plugin repository should include the #dsh topic tag. Plugins developed during the internal testing period can be migrated to the developer's personal account and made public. This means that DeepSeek is not only releasing the Harness product this time, but its accompanying plugin ecosystem will also be opened simultaneously. Currently, the official announcement has not been made, and specific details will be based on official information.

According to the Science and Technology Innovation Board Daily, the official version of DeepSeek V4 Pro has been updated to API, with the model name unchanged. The new version significantly enhances Agent capabilities, supporting Responses API and Codex integration. The API documentation shows that the official version of V4 Pro is priced at 3 yuan per million Token input and 6 yuan per million Token output, with cache hit input costing 0.025 yuan per million Token.Benchmark tests for the official version of DeepSeek V4 Pro show that its score in the terminal Agent test Terminal Bench is 87.9 points, approaching Fable 5's score of 88 points. In the AI safety agent evaluation benchmark CyberGym, as well as in the challenging agent evaluation benchmark AutomationBench, its performance even surpasses that of Fable 5.

DeepSeek's parent company, DeepSeek, completed the first round of signing for a new financing round on August 10 in Hangzhou. This round of financing has a scale of 50 billion yuan, corresponding to a pre-investment valuation of about 500 billion yuan, which is an increase of over 40% compared to the first round of financing completed in June (with a valuation of about 350 billion yuan). The earliest deadline for the first batch of fund delivery is August 30.This round of financing has a minimum investment threshold of 5 billion yuan, and the funds will mainly be used for computing power investment, model research, human resource expansion, and potential preparations for domestic listing. The fund entry paths are divided into two: part will directly enter DeepSeek's parent company, and part will enter a limited partnership controlled by founder Liang Wenfeng. Previously, DeepSeek briefly paused financing contacts at the end of July and resumed in early August. The new version of the model, DeepSeek-V4-Flash, was officially launched on July 31. According to OpenRouter data, the model reached a call volume of 72.2 trillion tokens in its first week, topping the global charts. The company has recently announced plans to raise API pricing and introduce a peak and off-peak pricing mechanism.

DeepSeek has just participated in a strategic placement of YuTree Technology with approximately 141 million yuan. The head of the Harness team, Cui Tianyi, subsequently publicly criticized the description of DeepSeek in the related IPO materials.The materials claim that DeepSeek V4 "ranks among the top tier globally in performance, with programming capability only 0.3% weaker than Claude flagship." Cui Tianyi pointed out that in the V4 paper, only one SWE Verified score is close to Opus 4.6, with scores of 80.6 and 80.8 respectively, which cannot be directly generalized to mean that the overall programming capability is only 0.3% weaker.He sarcastically stated that such writing is hard not to be considered "pure fabrication." However, the criticism was aimed at the exaggerated statements in the intermediary materials.

According to data from the multi-model aggregation platform OpenRouter, from July 27 to August 2, DeepSeek V4 Flash topped the global AI large model usage ranking with a call volume of 7.22 trillion tokens. Data from the open-source platform OpenCode shows that this model alone processed 8 trillion tokens on August 1. Four out of the top five positions are occupied by Chinese large models: Xiaomi's MiMo-V2.5 (5.1 trillion), Tencent's Hunyuan Hy3 (5.01 trillion), and another DeepSeek V4 Flash model (3.45 trillion) ranked second to fourth, while OpenAI GPT-5.6 Luna (2.99 trillion, a month-on-month increase of 738%) ranked fifth.The total call volume of Chinese large models last week reached 28.13 trillion tokens, surpassing the United States for 14 consecutive weeks, firmly holding the top position globally. The previously highly anticipated Kimi K3 fell out of the top ten this week, ranking 12th. It is noteworthy that free models are capturing an increasing share of traffic, with some models experiencing month-on-month growth exceeding 50%. Analysis indicates that domestic models have shifted from price competition to relying on inference performance and stable service capabilities, but companies like OpenAI and Anthropic continue to iterate, accelerating the internal competition and the pace of product iteration will determine user retention capabilities.

According to a report by a financial magazine, multiple trading insiders revealed that the large model company DeepSeek has restarted its second round of financing, planning to raise 50 billion yuan, with a pre-investment valuation of approximately 500 billion yuan, and plans to complete the signing by late August.Several investors mentioned that DeepSeek's second round of financing had at least started in mid-July, but was suddenly paused at the end of July. At that time, some investors on the negotiation standby list were informed that the plan to sign the financing agreement was postponed. According to media reports on July 26, DeepSeek paused its second round of financing, one reason being that DeepSeek's founder Liang Wenfeng was dissatisfied with widely circulated comments online, which mainly revolved around the suspected leak of the "transcript of the investor meeting."DeepSeek and the investors at the negotiation table hope that this financing can proceed discreetly after the restart. Currently, some investment institutions that had previously been actively in contact have stated that they have not received news of DeepSeek restarting the financing, and the channel remains in a paused state.

According to IT Home, the national supercomputing internet has officially launched the DeepSeek-V4-Flash-0731 model API calling service and model files.This version has undergone extensive post-training, significantly enhancing the agent's instruction-following capabilities, with performance in multiple benchmark tests comparable to the strongest closed-source models. Enterprises and developers can access it with one click without the need for environment configuration.The platform's AI community has gathered over 1,700 open-source large models and has become the first national resource pool of super-intelligent integrated computing power at the level of 100,000 cards.

DeepSeek releases update log, DeepSeek V4 Flash official version API is now live for public testing. The model structure, size, and DeepSeek V4 Flash preview of DeepSeek V4 Flash 0731 remain consistent, with only post-training being redone.The official version of DeepSeek V4 Pro will be released as soon as possible.

According to Bloomberg citing informed sources, DeepSeek plans to build a 1GW AI data center in Ulanqab, Inner Mongolia. The company will construct its own machine room while also renting computing power from other enterprises.Some capacity is expected to go online by the end of 2027 or early 2028. It is still unclear which type of chips the project will use.The 1GW scale will exceed the currently operational AI data centers of Chinese companies, but it is still lower than the 3GW and 5GW projects planned by American companies. DeepSeek is known for saving training computing power and is now also starting to build heavy infrastructure.

According to X platform blogger @MaxForAI, a screenshot of the DeepSeek Harness internal testing recruitment is circulating on social media. The screenshot shows that the DeepSeek Harness plans to start internal testing later this week, with the first batch of users being selected from small group chats. Selected individuals must submit personal information and sign a "Confidentiality Commitment Letter." The screenshot also specifically reminds that any leaks will result in disqualification and affect future opportunities for selection in the DeepSeek model and product internal testing. If the news is true, the official version of DeepSeek V4 and the Harness product are expected to be launched soon.

According to Bloomberg, DeepSeek announced the suspension of its second round of financing and verbally notified some potential investors that it would not sign investment agreements as expected. This suspension partly stems from founder Liang Wenfeng's dissatisfaction with the leak of a 4-hour internal transcript during its first financing deal, where he made 52 statements in an investor meeting regarding DeepSeek's organizational culture, open-source logic, technology roadmap, and views on the competitive landscape of the industry. The nearly 4-hour meeting record recently went viral on social media, including statements such as "The biggest gap between China and the U.S. in AI is computing power resources." Both DeepSeek and Liang Wenfeng have not confirmed the authenticity of the leaked record.DeepSeek stated that it may still choose to resume the transaction process in the future and has begun preparations for an initial public offering, with the IPO potentially being filed as early as this year.DeepSeek completed its first round of external financing in June 2026, raising over 50 billion yuan, approximately 7.4 billion dollars, with a post-investment valuation of about 50-59 billion dollars. It adopted a special structure, with most funds entering a limited partnership managed by Liang Wenfeng, where investors have no voting rights and are locked in for 5 years. Liang Wenfeng himself contributed about 20 billion yuan as the largest contributor, with participation from Tencent, CATL, JD.com, NetEase, and the Chinese National AI Fund. DeepSeek subsequently initiated discussions for its second round of financing in mid-July, targeting a valuation of about 71 billion dollars.

In a leaked internal conversation from the DeepSeek investor conference, DeepSeek founder Liang Wenfeng elaborated on his core views regarding the development of the domestic AI computing power industry and ecosystem. He stated that the current domestic AI chip replacement is facing a historic opportunity, with no compatibility barriers in hardware and ecosystem, and that NVIDIA cannot block the replacement process. The only current issue is insufficient production capacity, and it is expected that within a year, the market's perception of domestic chips as "unusable and poor-performing" can be reversed.He pointed out that NVIDIA's CUDA ecosystem moat is rapidly collapsing, and multiple industry transformations combined with export controls have opened a development window for domestic AI chips. He revealed that DeepSeek has partnered with Huawei to obtain 16,000 950 cards to run the ecosystem, and Huawei's 950 super nodes can completely replace NVIDIA's GB200 and GB300 in terms of performance and price, showing an optimistic attitude towards the development of the domestic computing power industry.

Last month, comments from Liang Wenfeng during the DeepSeek investor meeting have recently circulated. Liang Wenfeng clearly stated that DeepSeek's goal is AGI, rather than becoming the next super app. They will not engage in 3D, video generation, or world modeling, nor will they close source code, stating, "Restraint is a strategy used to increase the probability of achieving AGI." He believes that products are merely "by-products" on the path to AGI, with the current focus on Coding Agent and continuous learning capabilities. The next phase aims to achieve autonomous model iteration, with the endpoint pointing towards embodied intelligence.In terms of commercialization, Liang Wenfeng stated that they only seek to earn "reasonable profits" and do not set profit maximization as their pricing goal. DeepSeek is "still far from a complete shift to commercialization." He assesses that the gap between AI in China and the U.S. mainly lies in resources rather than talent, stating, "Using a fraction of the computing power can narrow the gap to 6 months, 3 months." Regarding industry competition, he believes that open source is "the sweet point for companies of this scale," while also predicting that domestic large models will ultimately converge to "two large companies and two small companies." Liang Wenfeng also mentioned that maintaining team stability is the only non-negotiable core principle, and the company is driven by vision, with "a group of ordinary people achieving extraordinary things."

According to PConline Pacific Technology reports, the official version of DeepSeek-V4 has begun a small-scale gray testing, including two versions: DeepSeek-V4-Pro and DeepSeek-V4-Flash. Insiders say that the official version is expected to be released at the end of this month, which is somewhat delayed compared to previous plans. The new version will focus on optimizing inference efficiency, response speed, and agent capabilities.Developers participating in the testing claim that V4's comprehensive capabilities are on par with Anthropic Opus 4.8, with code generation close to OpenAI GPT-5.6 Sol, and improvements in scenarios such as 3D modeling, SVG image generation, and mini-game development. DeepSeek-V4 has completed the adaptation for de-CUDA, and on the first day of launch, it will be compatible with multiple domestic AI chips, and will adopt peak and off-peak API billing, with peak prices being double that of off-peak.