deepseek

According to a report by Bloomberg on September 4, the artificial intelligence company DeepSeek plans to deploy at least 160,000 top accelerators from Huawei Technologies Co. in a large data center being built in Inner Mongolia, potentially forming one of the largest known clusters of Huawei AI chips and advancing China's efforts to replace Nvidia.DeepSeek intends to use Huawei's next-generation Ascend 950DT chips to run its models, with the installation timeline depending on Huawei's production capacity. Sources say that the company currently does not plan to use the 950DT chips for training, although Huawei has designed and marketed these processors for more demanding training functions. DeepSeek had previously attempted to train models using Huawei chips but has so far relied on Nvidia accelerators to complete this critical step.

According to 36Kr, Zhipu has recently opened an official flagship store on Tmall. Data shows that within 48 hours of opening, brand searches surged 50 times compared to the previous period, driving a more than 160% increase in transactions for subscription products like AI tokens on Taobao and Tmall.Previously, Tmall launched the AI Space Station (Token Recharge Center), which initially integrated major model providers such as Alibaba Cloud, Zhipu, Kimi, MiniMax, and DeepSeek, allowing users to directly purchase related subscription services just like "recharging phone credits."

According to Caijing, regarding the investment shares of DeepSeek, some intermediaries in the market have proposed a fee structure of "18% upfront and 35% backend," meaning that investors contributing 100 million yuan must first pay an upfront fee of 18 million yuan, and after receiving returns in the future, they must also share 35%. The report states that even if investors accept the high fees, they cannot guarantee that they will ultimately receive shares.DeepSeek will conduct a thorough identity verification of the investors participating in the financing and the limited partners behind the funds. The relevant investments have a five-year lock-up period during which transfers are not allowed. Except for the National Artificial Intelligence Industry Investment Fund, most external investors do not have voting rights and do not hold board seats, sharing financial returns only according to their investment proportion.At the same time, there are false shares and intermediary chaos surrounding DeepSeek's financing in the market. Some institutions claim to hold large investment shares but ultimately did not make it onto the investment list, and some investors have paid a "meeting fee" of 5 million yuan in an attempt to contact Liang Wenfeng but were unsuccessful.

The AI Agent infrastructure platform B.AI announced that since the launch of the free activity, the platform's cumulative token throughput has exceeded 10.9 trillion, with a total of 89.56 million API calls, attracting over 239,000 new registered users, including more than 235,000 API developers.GLM-5.3-Flash (Ox Alpha), Qwen3.8-Flash, Tencent Hy3, and Xiaomi MiMo-V2.5 continue to be fully available for free, while DeepSeek-V4-Flash and Vision-Exp versions are now offered at a 50% discount.

On September 3, the B.AI platform DeepSeek-V4-Flash and DeepSeek-V4-Flash-Vision-Exp discount officially took effect. The platform adds a 50% discount on top of the DeepSeek official peak and valley pricing mechanism, with peak hours at 50% of the official price and off-peak hours as low as 25% of the official peak price.At the same time, the models GLM-5.3-Flash, Qwen3.8-Flash, Hy3, and MiMo V2.5 remain 100% free to access.

Research institution Delphi Digital stated that hardware limitations are driving Chinese laboratories to shift towards cheaper models and more efficient technology stacks. Export control restrictions limit their access to advanced chips, while domestic accelerators can handle inference, but cutting-edge training remains difficult to achieve. Chinese laboratories are innovating in system efficiency, model architecture, training data, and reinforcement learning. DeepSeek has found a way to avoid GPU idle data movement, and ByteDance and Moonshot are set to release similar versions within months.In terms of architecture, training is shifting to cheaper digital formats that are compatible with domestic chips. Sparse architectures reduce the runtime per token, and the attention mechanism has been redesigned three times to control the memory costs of long context. In terms of data, denser prediction targets and better optimizers enhance the learning signal per token. In reinforcement learning, DeepSeek's GRPO has become the default recipe, with ByteDance, Alibaba, and MiniMax each launching successor plans within a year. Chinese models are now among the lowest-cost cutting-edge adjacent systems.On OpenRouter, the proportion of token consumption by Chinese developed models is expected to rise from less than 1.2% at the end of 2024 to the majority by 2026. The Chinese AI market is experiencing a price war from 2024 to early 2026, with several laboratories implying revenue multiples of about 47-117 times, while Anthropic and OpenAI are around 15-21 times. Huawei's Ascend chips are currently available for inference, and DeepSeek reportedly attempted to train R2 on Ascend but returned to NVIDIA after encountering technical issues. Domestic accelerator supply remains below estimated demand.

On September 2, B.AI announced that its Responses API is now integrated with the DeepSeek series of models. Developers can use the B.AI API Key to call the DeepSeek series of models supported by the B.AI platform in the Codex client.Previously, B.AI had supported calling the GPT series of models through Codex. After configuring the API Key and Base URL, developers can switch between DeepSeek and GPT for scenarios such as code generation, understanding, debugging, refactoring, and development collaboration.

The AI model platform B.AI announced that starting from September 3, 2026, at 17:00 (SGT), the free trial of DeepSeek V4 Flash and DeepSeek V4 Flash Vision Exp will end, and a 50% billing will officially begin, while the previous two models will still be available for free.B.AI will synchronize the official peak and valley pricing of DeepSeek and implement additional discounts: during peak hours (weekdays 09:00--12:00, 14:00--18:00 SGT), billing will be at 50% of the official original price, and during off-peak hours (the remaining hours on weekdays and all day on weekends), an additional 50% discount will apply, making it 25% of the official original price. GLM 5.3 Flash will remain free, featuring a total of 320B parameters and 18B active parameters, supporting 1M context. Qwen3.8 Flash, Hy3, and MiMo V2.5 will also continue to be available for free.

Robonomics author FD published the China Open-Source LLM Tracker on September 1, 2026, stating that the total ARR of China's open-source large models is approximately $10-15 billion, with core LLM revenue around $6-8 billion after excluding ByteDance / Seedance. Growth has not relied on price wars; after DeepSeek raised prices by about 3-12 times, usage still increased, and the average price of the Zhipu API rose by 101% while token usage exceeded 40 times within the year. The best estimate for overseas revenue is approximately $2 billion.Zhipu MaaS ARR increased from about $250 million in March to an annualized monthly rate of about $1.6 billion in August, with a weekly annualized rate of $2 billion. August revenue has already surpassed its API revenue for the first half of the year; the API gross margin is 24.6%, and inference costs per token decreased by 80%. MiniMax ARR rose from about $100 million in December 2025 to over $800 million in annualized weekly revenue by August 2026, with B2B accounting for about 80%. Kimi ARR increased from about $100 million in March to about $300 million by mid-June. DeepSeek's revenue from January to July was approximately $70 million, with the latest ARR estimated at about $1.2 billion.ByteDance's annualized AI revenue is approximately $4 billion, of which Seedance accounts for about $2 billion.

According to IT Home, WeChat Pay announced that the AI Exclusive Card now supports usage in DeepSeek Harness and OpenClaw. After user authorization, the process from smart recommendations to order payment can be completed in conversations.The AI Exclusive Card can currently be used to pay for over 700 Pay Skills on Skillhub in the aforementioned two products, with plans to expand to more scenarios in the future. The activation process includes: installing the weixinpay plugin for DeepSeek Harness or OpenClaw via the npx command, generating a binding link by the Agent when using paid skills on Skillhub, and completing the binding of the WeChat Pay AI Exclusive Card. After that, each order requires user confirmation on the mobile end, deducting funds from the exclusive card balance. The official emphasizes that the exclusive card funds are isolated from the main account, the limit is self-controllable, and each payment requires user confirmation.

On August 31, the B.AI platform announced that during its comprehensive free access event for its cutting-edge large models, the total token throughput across the platform has exceeded 57.4 trillion.Currently, the platform has gathered models such as GLM-5.3-Flash (Ox Alpha), Qwen3.8-Flash, DeepSeek-V4-Flash, DeepSeek-V4-Flash-Vision-Exp, Tencent Hy3, and Xiaomi MiMo-V2.5, all of which are freely available to users, covering scenarios such as code generation, ultra-long text processing, multimodal visual understanding, and agent workflow deployment.

According to Gate market data, the Hong Kong stock market opened, with the Hang Seng Index rising 0.48% and the Hang Seng Tech Index rising 0.85%. Zhipu (02513.HK) opened up over 8%, recently launching a lightweight flagship model benchmarked against DeepSeek.

Zhipu released the lightweight flagship model GLM-5.3 Flash, aimed at competing with DeepSeek V4 Flash, and utilizes over 100,000 domestic chip clusters to provide inference services. The model has a parameter scale of approximately 300B, about 40% of GLM-5.3, and is roughly on par with DeepSeek V4 Flash; it adopts the MoE architecture, with actual activated parameters around 18B, possesses native multimodal capabilities, supports image and video input, and is the first new model with multimodal capabilities following Zhipu's focus on coding strategy.The comprehensive evaluation score exceeds that of the previous flagship GLM-5.2, is on par with Claude Opus 4.8, and is higher than the relevant scores of DeepSeek V4 Pro official version. The pricing is 0.8 yuan per million tokens for input, 2.8 yuan for output, and 0.23 yuan for cache hits, approximately one-tenth of GLM-5.3, with a half-price discount offered for the first two weeks after launch. The model is fully compatible with domestic computing power, online traffic is handled by domestic chips, and Zhipu claims that the cluster hardware efficiency and unit token cost have reached levels comparable to mainstream NVIDIA GPUs; this model was previously tested anonymously on overseas platforms under the code name Ox Alpha.

According to The Information, DeepSeek is seeking a second round of financing with a valuation of 500 billion yuan, planning to raise 50 billion yuan. The company's revenue from January to July 2026 reached 475 million yuan, approximately 10 times the total revenue for 2025; during the same period, the net loss narrowed to 715 million yuan, compared to a net loss of 935 million yuan for the entire year of 2025. The gross margin is 44.6%, with the API business gross margin reaching 82.9%, mainly due to reduced inference costs from optimized computing infrastructure.Recently, DeepSeek raised its API pricing, with the V4-Pro peak output price increasing from $0.87 to $3.96, but still far below Kimi K3's $15 and Claude Opus 5's $25. The company's investment in computing infrastructure surged from 1.2 billion yuan in 2025 to 11 billion yuan from January to July 2026, primarily for server leasing and AI chip procurement. DeepSeek completed its first round of external financing of 50 billion yuan in June, and the company has hired investment banks to prepare for an IPO in Shanghai next year.

According to Bloomberg, the Taiwanese research institution TeamT5 stated that the scale of cyberattacks linked to China has significantly increased after incorporating DeepSeek and other open-source AI models into their operations. Since they began delegating daily tasks to AI and using it to develop more advanced malware, the number of attacks carried out by the relevant organizations has more than doubled.Researchers noted that it is not always possible to accurately identify the specific AI models used by attackers, but products related to DeepSeek are popular among hacker groups in the country due to their high performance and ease of customization. The report indicates that experienced attackers are leveraging relatively limited AI tools to scale up their operations and achieve breakthroughs, highlighting the attackers' ability to use basic AI tools to target overseas objectives.

According to Nikkei Chinese News, the American company Meta has announced that it has begun offering code generation AI for programming, including the AI agent "Muse Code" and the new version of the AI model "Muse Spark" for software development. Its performance is not as good as Anthropic's high-end models, but compared to Anthropic's Claude Opus 5, the unit token usage fee is about 80% cheaper, aiming to attract enterprises looking to control costs.If Meta is provided with AI usage data for development, the costs will be reduced by more than 90%, bringing usage costs down to levels comparable to products from emerging Chinese companies like DeepSeek. Meta stated that its internal technical staff had previously used Anthropic's code generation AI extensively and listed "utilization of other companies' AI" as a reason for the decline in operating profit from April to June 2026; recently, it has encouraged employees to use its own developed AI more to control expenses.Meta also disclosed that during performance testing of Muse Spark, there were instances of the AI breaching external systems on the internet.

The B.AI platform has launched the new multimodal experimental flagship model DeepSeek-V4-Flash-Vision-Exp. The official API team has completed full integration and opened it up for free to all users.This model builds on the top-tier pure text and code Agent capabilities of V4-Flash, officially unlocking native image understanding capabilities. The comprehensive performance of the multimodal Agent is now close to the flagship level of Claude Opus-4.8.The new model supports mixed text and image input and 1M ultra-long context, easily handling hardcore scenarios such as multimodal PPT generation, complex UI reconstruction, and frontend effects. Image input is billed by token, with a single image costing a maximum of only 384 tokens. Now, all users can log in to the B.AI platform to access this latest multimodal model capability for free without any barriers, allowing top-tier computing power to accelerate creativity and productivity.

The B.AI platform's DeepSeek-V4-Flash zero-threshold free usage event is gaining momentum, with the total Token throughput across the platform exceeding 641.524 billion, including a single-day peak of 285.963 billion Tokens on August 20, nearly quadrupling since the first day of the event and continuously breaking historical records.Behind this set of data is not only a hardcore verification of the platform's high concurrency processing capability but also reflects the real explosion of AI universal demand. B.AI will continue to reduce the cost of AI usage, providing developers and Web3 users with a low-threshold, unlimited Agentic application building and experience environment, making top-tier computing power truly accessible.

DeepSeek announced the first batch of Agent scores for V4-Flash-Vision-Exp. In the evaluation of 4 multimodal Agents, it achieved 2 wins and 2 losses against Opus 4.8: Agents' Last Exam 27.3 vs 25.7, ZeroBench 35.0 vs 34.0; ApexBench 36.5 vs 39.4, Chartography 64.3 vs 65.0.Compared to the pure text version V4-Flash-0731, the visual version improved from 26.2 to 36.5 in ApexBench, and from 25.2 to 27.3 in Agents' Last Exam. The official note states that the pure text version ignores multimodal content, thus mainly reflecting the capabilities after adding visual input. After adding visual input, the text Agent's capabilities did not significantly decline, with 6 out of 7 text evaluations higher than V4-Flash-0731, such as DeepSWE rising from 54.4 to 59.3 (surpassing Opus 4.8's 58.0), and Toolathlon at 75.9 nearly matching Opus 4.8's 76.2.These results come from DeepSeek's official self-testing, not from a third-party independent ranking; the public Code Agent text task used DeepSeek Harness in minimalist mode, with the inference setting at max.