model

Theo báo cáo của Rest of World, Alibaba đã ra mắt nền tảng mô hình mã nguồn mở ModelScope vào năm 2022, OSChina đã ra mắt MoArk vào năm 2023, cung cấp dịch vụ lưu trữ, thử nghiệm và tùy chỉnh mô hình cho các nhà phát triển tiếng Trung trong tường lửa. Các cơ quan quản lý Trung Quốc đã chặn Hugging Face vào năm 2023 mà không công bố lý do cụ thể; báo cáo cho biết, chính phủ đồng thời ngầm cho phép các phòng thí nghiệm trí tuệ nhân tạo vượt qua lệnh cấm để chia sẻ mô hình Trung Quốc ra toàn cầu. ModelScope hiện đang lưu trữ hơn 170.000 mô hình và vào tháng 3 đã cho biết có 170.000 mô hình và 250 triệu người dùng; MoArk lưu trữ khoảng 20.000 mô hình phổ biến, cả hai vẫn thấp hơn so với Hugging Face, nơi lưu trữ hơn 3 triệu mô hình mở.Giám đốc điều hành OSChina, Xu Yong, cho biết không phải ai cũng có thể sử dụng VPN liên tục, Trung Quốc cần xây dựng một hệ sinh thái trí tuệ nhân tạo độc lập phục vụ người dùng tiếng Trung, và cho biết Trung Quốc đang phát triển một hệ sinh thái độc lập nhanh hơn cả thời kỳ Internet. Ông cũng cho biết, MoArk đã triển khai đội ngũ kỹ sư để đảm bảo mô hình có thể chạy trên các loại chip Trung Quốc khác nhau, đảm bảo rằng việc thích ứng các mô hình mở chính thống với chip Trung Quốc là sứ mệnh của họ. Hai nền tảng này cung cấp một số dịch vụ miễn phí, tính phí cho sức mạnh tính toán bổ sung và các tính năng cao cấp; ModelScope còn tổ chức hackathon đại học và mở không gian làm việc chung và sự kiện cho các doanh nhân trí tuệ nhân tạo tại Hàng Châu.Hugging Face có trụ sở tại New York là nền tảng chính để chia sẻ mô hình và tập dữ liệu mở. Vào tháng 9, NVIDIA đã thông báo mua lại công ty này với giá 12,9 tỷ USD và trong tài liệu quản lý đã chỉ ra rằng các cơ quan quản lý có thể cấm chia sẻ mô hình Trung Quốc trên trang web này.

Theo báo cáo của Bloomberg, công ty khởi nghiệp Model ML cung cấp dịch vụ phần mềm AI cho các tổ chức tài chính như ngân hàng đầu tư đang đàm phán huy động vốn hơn 100 triệu USD, với định giá vượt quá 1 tỷ USD. Hiện tại, vòng huy động vốn này vẫn chưa hoàn tất, các chi tiết cụ thể có thể thay đổi.Các nguồn tin cho biết, quy mô huy động vốn có thể mở rộng lên 150 triệu USD tùy thuộc vào sự quan tâm của các nhà đầu tư.

ChainCatcher thông báo, API B.AI chính thức ra mắt "nhóm giảm giá 10% chính thức" đầu tiên, trong khi đảm bảo tính ổn định của việc gọi mô hình, chất lượng dịch vụ và hiệu quả chi phí, dựa vào kênh kết nối trực tiếp chính thức, cung cấp dịch vụ gọi AI đáng tin cậy cho người dùng. Lô đầu tiên đã hỗ trợ các mô hình trong chuỗi Claude, nhiều mô hình chính khác sẽ lần lượt được mở. Đồng thời, mô hình "nhà cung cấp dịch vụ tự chọn" của nền tảng có thể được hưởng mức giảm giá tối thiểu 20%, trao quyền tự chủ đầy đủ cho người dùng, có thể linh hoạt cân nhắc hiệu suất, tính ổn định và chi phí dựa trên nhu cầu thực tế của tình huống kinh doanh, đạt được cấu hình tối ưu.

The "Anthropic" plan allows the EU's cybersecurity agency to use the powerful artificial intelligence tool "Mythos," which can discover and exploit vulnerabilities in computer systems.According to informed sources, Anthropic will allow the European Cybersecurity Agency to join the "Glass Wings" project. This project aims to let key institutions test the capabilities of "Mythos" before widespread deployment. A spokesperson for the European Commission stated that negotiations with Anthropic are progressing but declined to comment further.

The AI service platform B.AI announced that it has completed the underlying access to MiniMax's next-generation multimodal large model M3, and it is now available for API and web access.It is reported that the MiniMax M3 model is based on the MSA sparse attention mechanism, natively supports multimodal understanding, code generation, and parsing, and has a 1M context window and long text reasoning capabilities, covering various application scenarios from dialogue to professional development. Currently, developers and users can directly access and use this model through the B.AI platform.

MINIMAX-W (00100.HK) announced the launch of the MiniMaxM3 model, stating that "MiniMaxM3 has achieved cutting-edge capabilities in specialized tasks such as programming and agents. It utilizes a newly proposed attention architecture MSA (MiniMaxSparseAttention), supporting up to 1M ultra-long context. As expected by the outside world, it is also a native multimodal model, supporting input from images and videos, and can operate a computer desktop.These three capabilities are essential for overseas closed-source cutting-edge models. M3 is the first domestic model to possess these elements and is currently the only open-source model."

AI model aggregation platform OpenRouter announced the completion of a $113 million Series B funding round, led by CapitalG, the growth fund of Google's parent company Alphabet. Participants in this round of financing include a16z, Menlo Ventures, NVentures under NVIDIA, as well as ServiceNow, MongoDB, Snowflake, and Databricks, with the company's valuation exceeding $1 billion.

10x Research stated that its Bitcoin trend model has turned bearish. It pointed out that the market narrative of Strategy (formerly MicroStrategy) "never selling Bitcoin" has begun to waver. 10x Research mentioned that Michael Saylor recently hinted that the company might sell part of its holdings of 843,000 BTC in the future, and this change has led to approximately $2.7 billion in outflows from related Bitcoin ETFs.

According to Jinshi News, the technology company Hangzhou Yunshen Technology Co., Ltd., which focuses on embodied intelligent robots and their supporting products and solutions, has released a prospectus. It shows that the revenue from the embodied intelligent robot project for the year 2025 is expected to reach 322.3773 million yuan; the revenue from product components is expected to reach 11.9501 million yuan; and the revenue from supporting services is expected to reach 2.573 million yuan.According to the resolutions from the company's first extraordinary general meeting of shareholders in 2026 and the ninth meeting of the first board of directors, this issuance aims to raise 2,502.511 million yuan. After deducting issuance costs, the funds will be invested in the research and development projects for embodied algorithms and models, the research and development projects for robot bodies and solutions, the industrialization projects for embodied intelligent robots, and the construction projects for embodied intelligent robot bases.

The ALEX Lab Foundation submitted governance proposal AGP-8, which plans to make structural adjustments to the ALEX protocol, including stopping the emission of ALEX community tokens, closing the treasury grant program TGP, and introducing a protocol-driven token buyback and burn mechanism. Currently, the circulating supply of ALEX is approximately 973 million tokens, close to the cap of 1 billion tokens.If the proposal is approved, the next 32 cycles will be the final ALEX emission cycles, after which there will be no new token emissions. Approximately 1.568 million STX remain unclaimed in the TGP 2024 treasury, and after a 30-day grace period, the ALEX Lab Foundation will use these funds to buy back and burn ALEX tokens at market price. Future protocol revenue will also be used for ongoing buybacks and burns after covering operational costs. This proposal aims to shift ALEX from an inflationary model to a deflationary model.

Alibaba (09988.HK) mentioned in its financial report that it is continuously expanding the boundaries of AI capabilities through deep innovation. Recently, by launching a series of brand new large language models and multimodal models, we have made significant progress in model intelligence. In March, we released Qwen3.6-Plus, achieving a remarkable overall performance improvement, particularly excelling in programming and agent programming, becoming an industry benchmark in front-end web development and complex warehouse-level tasks.Qwen3.6-Plus also possesses stronger multimodal perception and reasoning capabilities, with a native context window length supporting up to 1 million tokens, further enhancing stability and reliability. In addition to the Qianwen model family, we are also continuously enriching our specialized model layout, including the world model HappyOyster that supports real-time creation and interaction, as well as the multimodal model HappyHorse for video generation, both of which are currently undergoing gray testing.

The cryptocurrency research institution Delphi Digital released the latest report "How Far Can Saylor Stretch It," which systematically analyzes the Bitcoin (BTC) funding expansion mechanism of Strategy, pointing out that its financing structure is transitioning from "low-cost accumulation" to the "diminishing marginal efficiency" stage. The report shows that in the current asset accumulation system centered around Bitcoin, STRC has become the core financing tool for Strategy's continuous purchase of BTC. Initially, it relied on a significant premium in MSTR's stock price (mNAV far exceeding BTC's net value) to achieve a positive cycle of "issuance leads to accumulation," but as the valuation has fallen back to about 1.24 times the EV-based mNAV, the BTC per share enhancement effect from common stock issuance is nearing breakeven.At the same time, while convertible bond tools have played an important role historically, they have accumulated about $8.2 billion in principal and will face concentrated repayment pressure after September 2027, putting long-term sustainability of the financing structure under pressure. STRC provides a continuous financing source for Strategy by offering approximately 11.5% annualized monthly dividends to income-oriented investors, to maintain the pace of BTC purchases. However, this mechanism also introduces ongoing cash flow obligations, meaning that each round of financing increases BTC assets while simultaneously accumulating future dividend burdens.The report emphasizes key risk scenarios: if BTC prices remain stagnant and MSTR's premium fails to recover, then the "STRC financing purchase gain" may be gradually offset by "common stock dilution and dividend obligations." Although the company's approximately $2.25 billion cash reserves can cover about $1 billion in redemption pressure in 2027, larger-scale debt and dividend structures in 2028 still need to be addressed. Additionally, the current authorized issuance limit of about $28.3 billion for STRC becomes a critical constraint point. Once the limit is reached, the ability to purchase new BTC may slow down, but existing dividend obligations will continue to exist, thus altering the overall BTC per share dynamic growth path.

Binance released its latest security report, addressing the current industry situation of rapidly spreading AI scams. The platform has deployed over 24 AI security programs and equipped more than 100 AI models to build an intelligent defense system against various types of cryptocurrency fraud.Statistics show that from early 2025 to the first quarter of 2026, Binance has protected over 5.4 million users and intercepted potential fund losses of $10.53 billion. In Q1 2026, the platform successfully intercepted 22.9 million scam and phishing attacks, protecting user funds amounting to $1.98 billion, with an average of over 9,600 real-time risk alerts pushed daily, and a total of 36,000 malicious on-chain addresses blacklisted.The report pointed out that deepfakes, voice cloning, and phishing bots have become mainstream scam tactics, with the overall scale of cryptocurrency fraud reaching $17 billion in 2025, a year-on-year increase of 30%. In terms of risk control, Binance's AI system handles 57% of fraud detection work, reducing the credit card fraud rate to 60%-70% of the industry average; it has upgraded AI anti-counterfeiting KYC reviews, with review efficiency improved by up to 100 times.The AI trading tool Binance Ai Pro uses an isolated account structure, only allowing trading permissions and prohibiting withdrawals, with the platform intercepting 12% of high-risk third-party AI plugins. Additionally, in 2025, Binance assisted in recovering $12.8 million in scammed funds, handled 48,000 cases, and collaborated with law enforcement to freeze $131 million in illegal assets.

According to GMGN monitoring, the Base ecosystem token POD (dphn.ai) saw a significant increase in market value this morning, peaking over $170 million, currently reported at $130 million, with a 24-hour increase narrowing to 82%, and a trading volume of $6.3 million during the same period.It is reported that dphn.ai is an AI model development and DePIN project, and the "Dolphin Mistral 24B Venice Edition" it provides is the default uncensored AI model on the Venice platform. Previously, during the surge of VVV, POD rose as it served as a core model provider. This increase may also be compounded by the influence of Base co-founder Jesse's retweet, with related sentiment continuing to ferment.

According to GMGN monitoring, the Base ecological token POD (Dolphin) surged significantly this morning, with a peak market value exceeding $82 million, currently falling back to $76.4 million, with a 24-hour increase of 130%, and an on-chain trading volume of approximately $2.9 million during the same period.The catalyst for this round of increase may be the strong market performance of VVV. As an AI model development and DePIN project, Dolphin provides the "Dolphin Mistral 24B Venice Edition," which is the default uncensored AI model on the Venice platform, leading to an increase in POD as the core model provider.

According to CoinDesk, European Central Bank President Christine Lagarde stated in a speech that Europe should not simply replicate the U.S. stablecoin model, but should prioritize building tokenized settlement infrastructure centered around central bank currencies.She pointed out that the $310 billion stablecoin market, dominated by Tether and Circle, poses financial stability risks and could transmit pressure to underlying asset markets during times of crisis. Lagarde believes that the technological advantages of stablecoins can be replicated by central bank infrastructure, while their monetary functions introduce unacceptable financial stability risks. She cited the 2023 collapse of Silicon Valley Bank, which led to a brief de-pegging of USDC, as an example.Although a consortium of 12 major European banks, including ING, BBVA, and BNP Paribas, plans to launch a private digital euro later this year, citing that Europe faces "dollarization risks," Lagarde still called for EU countries to support the research and development of central bank digital currencies. The European Central Bank aims to launch a digital euro by 2029.