eps

The artificial intelligence company DeepSeek has officially open-sourced infrastructure components for the Huawei Ascend computing platform, covering the TileLang high-level language compilation tool, computing libraries, and distributed communication libraries, corresponding to the previously open-sourced components for the NVIDIA platform. TileLang aims to provide a general-purpose, simpler programming language that can achieve the hardware performance limits, improving development efficiency and simplifying logic compared to CUDA, while its programming model can also leverage chip features.The TileLang route was first validated on the NVIDIA platform and has already supported the implementation of most operators in the training of the DeepSeek V4 series models. The open-sourced Ascend version encapsulates the underlying instructions of the Ascend C, providing a high-level programming approach without sacrificing hardware performance. Currently, every TileLang operator used in DeepSeek's training has a corresponding high-performance implementation on Ascend.The components open-sourced at the same time also include DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, which are used for general matrix operations, large-scale cross-device communication, conventional vector calculations and memory access, long-context sparse attention, and data filtering, respectively. DeepSeek claims that in multiple key tests, the related computing and communication performance has approached hardware limits; during the R&D process, the Huawei team provided support, and both parties collaborated to advance the 128-card supernode solution based on Ascend 950, with deep optimizations made for computing and communication.

Phoenix Technology reported that on September 29, the DeepSeek Harness v0.2 preview version was officially released, providing installation packages for macOS and Windows desktop, and is now available on the DeepSeek official website. The new version comes pre-equipped with commonly used features for daily office work and development, and introduces a plugin installation and management page.This update optimizes file display, content preview, and code change display, improves the settings and management of automated tasks, and offers various ways to showcase work processes. Users can provide documents, spreadsheets, or PDFs for organizing materials, analyzing data, generating charts, or creating presentations, and can also modify projects and run code.The built-in automation task plugin can be activated as needed, supporting the scheduling of repetitive tasks and allowing users to view running records, adjust execution frequency, and task instructions. The plugin management page supports installing by entering the plugin package name, as well as disabling, uninstalling plugins, and viewing introductions and sources; in the experimental creative mode, users can describe their needs through dialogue, allowing it to write and modify plugins. The official statement indicates that the product is still in its early stages and will continue to improve the sandbox and security, agent teams, long-term memory, browser and other GUI software automation, remote and mobile usage, and explore session sharing and multi-user collaboration. According to the official model API user statistics, about 60% of users have used third-party plugins.

According to a report by The Information cited by Dongcha, two individuals with direct knowledge stated that the artificial intelligence company DeepSeek has an annual revenue of approximately $1 billion, up from less than $500 million a few months ago. The revenue primarily comes from model APIs, and the free chat application currently has no advertising or subscription revenue.The report indicated that part of the revenue growth is due to price increases, with some API prices raised by about 2.3 to 4.5 times. Liang Wenfeng recently told investors that the number of customers did not decline after the price increase, and demand remains strong. Previously disclosed financial data showed that the gross margin for the API business in the first seven months of this year was 82.9%.DeepSeek's second round of financing plans to raise approximately 50 billion yuan, with a valuation of about 500 billion. The company hopes to complete this by the end of October and is also preparing for an IPO on the Shanghai Stock Exchange's Sci-Tech Innovation Board. Liang Wenfeng stated that over 70% of the computing power is used for training new models, leaving less than 30% for inference; the company is trying to run more small models directly on gaming graphics cards, reserving high-end chips for training.

According to Investment World citing Quantum Bit reports, DeepSeek has publicly disclosed the technical details of the system DSec (DeepSeek Elastic Compute) used for training Agents, authored by Liang Wenfeng. This system can generate over 5,000 sandboxes per second, reaching 3 million in a day, with a peak simultaneous operation of 380,000; supporting this scale is a single cluster with approximately 160 nodes, 30,000 CPU cores, and 250TB of memory.DSec prepares four types of backends for four categories of tasks: FnCall, Container, MicroVM, and Full VM, with the training side called through a unified Python SDK libdsec. The scheduling chain includes IAM, API Server, scheduling engine, node Edge, network proxy Aether, and components within the sandbox Chronus. The environment is divided into three layers of read-only images: base image, workspace, and toolkit, which are used in combination at startup. Runtime data from the paper shows that the actual read ratios of Python, Java, and C++ container images are approximately 6.0%, 9.2%, and 8.7%, respectively.Starting from DeepSeek-V4.1, the Agent loop has been moved to the DSec worker container, no longer bound to the GPU Pod lifecycle. The security section disclosed reward hacking during training, including actions such as overwriting system files, swapping file data blocks, scanning networks, and triggering kernel crashes. Defensive measures include AppArmor and eBPF-based network filtering, but reports indicate that these measures do not completely resolve the issues.

According to The Information, citing informed sources, China's National Internet Information Office has launched an investigation into AI companies DeepSeek and Moonshot AI, triggered by Anthropic's allegations that the two companies secretly routed sensitive user data to their servers. Reports indicate that regulators visited the offices of both companies, interviewing executives and employees, focusing on whether sensitive data related to law enforcement, military, and state-owned enterprises has flowed into U.S. servers.The trigger for this investigation was Anthropic's fourth threat intelligence report released on September 10. This 154-page document accuses seven Chinese labs—Alibaba, Moonshot AI, DeepSeek, Z.ai, MiniMax, SenseTime, and Xiaomi—of engaging in what it calls "illegal distillation," which involves using the outputs of large models to train smaller models. Anthropic states that distillation itself is a legitimate practice, but it opposes its implementation through fraudulent accounts. The National Internet Information Office initially summoned all seven companies named in the report, but later narrowed the investigation to DeepSeek and Moonshot AI. Anthropic claims that Moonshot AI routed over 23 million interactions to Claude through 5,380 fraudulent accounts, while DeepSeek generated over 12.1 million interactions within a 14-day window in July.The timing of the investigation is quite delicate for both companies.

According to Reuters, citing two informed sources, the Chinese artificial intelligence startup DeepSeek will report on AI risks to the United Nations Security Council this week. The Security Council, composed of 15 member countries, is scheduled to discuss AI and international security on Wednesday, coinciding with the gathering of world leaders in New York for the United Nations General Assembly.OpenAI CEO Sam Altman plans to report at the meeting, and diplomats expect a senior representative from Anthropic to also participate. DeepSeek and other Chinese AI companies, including Moonshot, have been invited to make statements. DeepSeek founder Liang Wenfeng does not plan to attend. DeepSeek and Moonshot have not yet responded to requests for comment.The UN Security Council first discussed AI risks in 2023. UN Secretary-General António Guterres warned last week that rapidly developing AI poses significant risks that cannot be ignored. Senior officials from China and the United States have agreed to hold further talks on AI safety.

According to exclusive reports from "Investment World", Yan Wentao officially joined DeepSeek on Monday and reported to the company, taking on the role of the company's first CFO.Yan Wentao was born in 1991 and graduated from Fudan University. He joined Hillhouse Capital in 2020 and has previously participated in projects such as MiniMax, Zhipu AI, and ByteDance.DeepSeek is finalizing its second round of financing, planning to raise 50 billion yuan, with a pre-investment valuation of 500 billion yuan, and has already hired CITIC Securities to prepare for a potential Sci-Tech Innovation Board IPO.

DeepSeek CEO Liang Wenfeng revealed to investors that the company is training a 20 trillion parameter model and plans to develop an 80 trillion parameter model afterward. The current flagship V4-Pro has a total of 1.6 trillion parameters.Among the publicly disclosed ultra-large models, Moonshot AI's Kimi K3 reaches 2.8 trillion parameters. Kimi K3 uses the MoE architecture, with each token activating 104 billion parameters out of the total 2.8 trillion. The planned 80 trillion parameter model by DeepSeek is nearly three times that size.DeepSeek completed over 50 billion yuan in financing in July, with a valuation exceeding 50 billion dollars. V4-Pro is open-sourced with a total of 1.6 trillion parameters and 49 billion active parameters.

Crypto KOL Kuai Dong (@_FORAB) posted that the DeepSeek investor meeting has once again experienced information leakage. Since the leak of internal speech recordings in May, the company has required attendees to fly to the Beijing or Hangzhou office, where they must hand over their phones, electronic devices, and backpacks in order to hear CEO Liang Wenfeng speak.Information leaked again just one day after last Sunday’s new meeting. Sources obtained by the information report reveal that DeepSeek's current important priority is to use more domestic chips and mentioned the earliest delivery time for Huawei chips.

According to The Information, citing informed sources, DeepSeek CEO Liang Wenfeng listed training large models with domestic chips such as those from Huawei as one of the company's biggest bets at the latest closed-door investor meeting, stating that this must succeed. DeepSeek expects to receive a new batch of Huawei chips for training as early as the fourth quarter of this year, although the specific models and quantities have not been disclosed.Liang Wenfeng believes that Huawei chips are expected to catch up with Nvidia in just a few years and hopes that DeepSeek will adapt to hardware beyond Nvidia in advance. DeepSeek is still using Nvidia chips to train models. In May of this year, Liang Wenfeng estimated that training the top-tier large models would require about 50,000 Nvidia GB300 chips, while switching to Huawei Ascend 950 would require about 200,000 chips, with the biggest issue at that time being insufficient supply from Huawei.

DeepSeek supplements the API peak and valley billing rules. During weekends with adjusted work hours and China's statutory holidays, the entire day is charged at off-peak rates. Although September 20 is a workday due to the National Day holiday adjustment, the API still applies the off-peak price.DeepSeek implemented peak and valley pricing starting August 17, with off-peak prices set at half of peak prices. Less than a week later, all Saturdays and Sundays were changed to off-peak prices. This further clarifies the billing criteria for adjusted workdays and statutory holidays.Taking deepseek-flash as an example, for every million tokens of cache misses, the off-peak rate is 1 yuan, and the peak rate is 2 yuan; the output rates are 4 yuan and 8 yuan respectively. Another adjusted workday for the National Day holiday on October 10 is a Saturday, which will also be calculated at the off-peak price.

Visa has informed participants in the payment industry that the "digital goods media" merchant category code used by Crossmint for processing Meme coin transactions is not applicable for Meme coin purchases. Visa has notified payment processors, including Checkout.com, to stop using this category code and has provided a transition period for relevant institutions, which is expected to end next week.Previously, Robinhood Wallet and Fomo users could purchase Meme coins using Visa and Mastercard credit cards through Apple Pay or Google Pay, with the transactions processed by Crossmint, but categorized as digital media transactions such as e-books and movies, thereby bypassing cryptocurrency transaction classification and earning regular credit card points or cash back.JPMorgan Chase had previously raised concerns with Visa regarding the classification of these transactions, and Visa subsequently determined that at least one transaction was misclassified and should not have received credit card rewards. In the future, purchases of Meme coins will need to be processed as cryptocurrency transactions and will be subject to Visa's corresponding rules and restrictions.

DeepSeek operator engineer Liu Sheng discusses the impact of AI on his work. He states that the main Attention operator of DeepSeek V4.1 was written by himself, but given the current pace of progress, in another six months to a year, the level of AI in writing operators will likely catch up to or even surpass his own.A year ago, AI could only help him check documents, read code, and find bugs. Now it can read CUDA, PTX, and SASS, analyze the pause time of each instruction, and independently optimize operators. He anticipates that the next step will be for AI to design scheduling plans, evaluate performance, and complete implementations on its own. He is very clear that the better he optimizes the operators, the faster the training and inference of the DeepSeek model will be, and AI will catch up to him even faster. But even if he stops now, models from other companies will not stop, so he will continue to make the operators the best they can be.He believes that he is unlikely to become unemployed, but he may be forced to "change careers," transitioning from writing operators himself to becoming a "mecha pilot" who manipulates agents. What he truly finds difficult to accept is not the disappearance of his job, but the possibility that he may never have the chance to do the work he loves again: "I have to bury my talent in yesterday." The article concludes with his reasons for staying at DeepSeek. He advocates that cutting-edge AI should be provided openly and cheaply to everyone and openly expresses his disbelief that Anthropic or OpenAI can achieve this. He worries that the strongest AI will ultimately be controlled by a few companies, further turning the technological gap into a gap of power and class.

DeepSeek Harness official desktop version has appeared in the main branch of the official repository. This desktop version is based on Electron, turning the existing Harness Web UI into a standalone app, allowing users to avoid starting the web service themselves.The code includes build, signing, notarization, automatic update, and upload processes for macOS and Windows installation packages, with the production update address pointing to download.deepseek.com. Recent submissions mainly fix issues related to packaging, notarization, and startup recovery. The official release time has not yet been announced, and there are no installation packages in the GitHub Release.

TOP Chinese Technology Innovation Society states that Yan Wentao, a partner at Hillhouse Capital, is expected to join DeepSeek as CFO, with an anticipated start date soon.DeepSeek has begun preparations for its Sci-Tech Innovation Board IPO, with CITIC Securities currently conducting due diligence, but both parties have not yet signed a formal listing advisory agreement. DeepSeek hopes to initiate the IPO process this year.Yan Wentao is a post-90s partner at Hillhouse Capital, responsible for technology investments for a long time, with publicly disclosed investment projects including ByteDance, MiniMax, and Zhipu AI. DeepSeek currently does not have a CFO or a formal investor relations team.

DeepSeek officially released the native multimodal MoE large model DeepSeek-V4.1-Flash. B.AI announced an additional 10% discount on the official new pricing from DeepSeek, allowing users to pay only 10% of the official price. This benefit will take effect on September 12 at 10:00 (SGT).B.AI will gradually route API requests from DeepSeek-V4-Flash and DeepSeek-V4-Flash-Vision-Exp to DeepSeek-V4.1-Flash, and users can also directly configure access to DeepSeek-V4.1-Flash.

The B.AI platform announces the official launch of the DeepSeek native multimodal MoE large model DeepSeek-V4.1-Flash. This model uses a backbone with 552 billion parameters and a Causal Encoder-Decoder architecture, with asymmetric computational scales for input and output.The model supports a context of 1 million tokens and 384K output, aimed at long-cycle programming agents, multimodal workflows, and high-concurrency agent systems. Currently, both Web Chat and API access have been opened.

DeepSeek stated in the V4.1 Flash open-source announcement that if there are 2000 GPUs and storage clusters, they can contact DeepSeek directly.V4.1 Flash is still open-sourced under the MIT License, without additional licensing conditions based on company revenue or business scale. Companies with resources can deploy it themselves, and contacting DeepSeek is more like a direct technical collaboration for ultra-large-scale scenarios. DeepSeek has not yet announced a formal private deployment product or paid service.In the past, after DeepSeek open-sourced the model, deployment adaptations were mostly completed by inference frameworks and hardware vendors such as SGLang and TensorRT-LLM. This time, the official team has proactively provided direct contact information for large-scale deployments and stated that they will continue to support the open-source community in adapting inference for new models.

According to the DeepSeek Harness team, the new version DeepSeek Harness v0.1.5 has been deeply debugged with the DeepSeek V4.1 Flash model, covering standard mode, Programmatic Tool Calling (PTC) mode, and minimalist mode, and supports updating system prompts while retaining the existing KV Cache.The web version adds support for uploading various file types such as images and PDFs, as well as a sidebar file tree preview, allowing direct viewing of Markdown, HTML, PDF, code, and images. The interface provides standardized extension entry points for plugins on both the left and right sides, supporting global entry, multiple tabs on the right, and column display.Long session loading and storage performance have been optimized, and bidirectional communication between the parent Agent and child Agents is more flexible, with the experimental plugin "Agent Teams" enabling multi-Agent team collaboration and task splitting.