Kimi K3正式开源
Related
Baidu's OCR model, Unlimited OCR, topped four charts: HuggingFace, GitHub, and more.
According to Mars Finance, Baidu recently officially released and open-sourced its end-to-end OCR model, Unlimited OCR. The model topped the GitHub Daily Trending and Python charts the day after its release, and also ranked first on HuggingFace's global overall model trending chart and multimodal model trending chart. Unlimited OCR is designed for long document parsing scenarios, with a total parameter size of 3B and only about 570M activation parameters during inference. Public evaluation results show that Unlimited OCR achieved a 93.92% overall score in the OmniDocBench v1.6 benchmark test, setting a new record for end-to-end OCR.
Of the approximately 3,000 unlicensed cryptocurrency exchanges in Europe, about 80% face the risk of being shut down.
The EU's MiCA (Mini-Asset Regulatory Framework) for the crypto asset market will officially come into effect on July 1st, requiring crypto service providers to obtain licenses or exit the market. Reports indicate that of the approximately 3,000 unlicensed crypto trading platforms in Europe, about 80% may face closure or cease providing services to local users due to their inability to meet the requirements by the deadline. This could potentially affect over 10 million users, who may be forced to migrate to compliant platforms.
MiniMax M3 has been officially open-sourced, with native multimodal support for millions of contexts.
According to Beating, a Chinese large-scale model manufacturer, MiniMax has officially open-sourced the weights of its native multimodal hybrid expert (MoE) model, MiniMax M3, on Hugging Face. MiniMax M3 boasts a total of 428 billion parameters, with 23 billion parameters activated by a single token, and natively supports over 1 million contexts. To reduce deployment memory overhead, the development team simultaneously released an MXFP8 quantized version, adapting it to mainstream inference frameworks such as SGLang, vLLM, and Transformers. In its multimodal design, MiniMax M3 performs joint training of text, images, and videos during the pre-training phase to achieve native semantic fusion, rather than performing multimodal alignment in the post-training phase. In terms of its operational mechanism, the model provides dual inference modes: a Thinking mode for complex logic and tool orchestration, and a Non-thinking mode for low-latency dialogue and code generation. The underlying kernel supporting over 1 million contexts is the simultaneously open-source lightweight attention kernel library, MiniMax Sparse Attention (MSA). Official data shows that MSA uses a grouped query attention (GQA) block retrieval mechanism. In a real-world test with an extremely long context of 1 million tokens, the MSA operator optimized for the NVIDIA Blackwell (SM100) architecture can achieve more than 9 times the pre-filling speedup and 15 times the decoding speedup compared to the traditional full attention mechanism, while significantly reducing inference overhead.
Google pressured the media to agree to license AI model training by threatening to stop paying Showcase licensing fees.
According to Beating's monitoring, Google has recently been promoting its News AI pilot program to news media with new conditions. Publishers participating in the pilot must license their content for AI model training; otherwise, they will no longer receive regular licensing fees after the old Showcase program is discontinued. Previously, Google paid media outlets a fixed annual fee through the Showcase program to display featured articles on pages like Google News and drive traffic to their websites. The new AI pilot program requires media to license their content for training large models to generate AI article summaries in Google News and Gemini. While Google claims the summaries are intended to encourage user clicks, with traffic from Google Search plummeting by nearly half, media outlets are generally worried that users will only obtain information from Google platforms, leading to a complete erosion of website traffic. Now, with little traffic remaining from Google, media attitudes have hardened significantly. Many executives believe that the remaining value from Google's traffic is no longer worthwhile for news content to train the Gemini robot, and some are even considering blocking Google's crawling altogether. To implement the new agreement, Google plans to gradually shut down the old Showcase program, turning annual fees into a bargaining chip for signing new projects. Media outlets that agree to AI licensing will continue to receive funding; otherwise, their collaboration fees will be completely cut off. The Washington Post and The Guardian have already joined the pilot program as initial partners. However, with Google's recent hardening stance, more media outlets are accelerating the establishment of independent subscription and registration systems, attempting to reduce their dependence on Google.
MetaPlanet's BTC reserves face a double-edged sword of exchange rate challenges following the Bank of Japan's interest rate hike; OSL Group, in conjunction with Hong Kong Polytechnic University, released a report stating that cross-border trade payments by enterprises will drive the large-scale adoption of stablecoins.
According to Mars Finance, BBX data shows that yesterday's interest rate hike in Japan and the Federal Reserve's decision window coincided, forming the most concentrated macroeconomic shock point this week. The latest research findings on stablecoins from institutions were also released on the same day. Key developments are as follows: Metaplanet Inc. (TSE: 3350), as Asia's largest corporate BTC reserve holder and the third largest Bitcoin holder among global listed companies (holding 40,177 BTC at an average price of approximately $104,000, with a target of 100,000 BTC by the end of 2026), faced double-edged sword pressure from exchange rates yesterday: After the Bank of Japan announced a 25 basis point increase in its policy rate to 1.0%, the strengthening of the yen superficially lowered the book value of BTC denominated in yen, but Bitcoin subsequently rose against the trend (CoinDesk's headline on the same day: "Bitcoin rallies after Japan rate increase"), resulting in an actual improvement in dollar-denominated holdings. The company currently holds approximately $2.64 billion (estimated at $65,750/BTC), and its latest capital action was the issuance of 8 billion yen (approximately $55 million) in its 20th EVO Fund on April 24th to purchase more BTC. Analysts warn that if the BOJ interest rate hike triggers a large-scale liquidation of "yen carry trades," global risk assets, including BTC, could suffer a systemic deleveraging shock—however, Metaplanet holds physical BTC rather than leveraged positions, so its main risk is the exchange rate conversion effect rather than forced liquidation. OSL Group (HKEX: |_2024111120230_) and CADI Business School of Hong Kong Polytechnic University jointly released a white paper on June 16 (recorded by The Block at 9:01 am EDT on the same day), titled "Cross-border Trade Payments Will Drive the Adoption of Regulated Corporate Stablecoins." The core argument is that enterprise-level cross-border trade payment demand is the main path to driving the large-scale adoption of regulated stablecoins, and its importance surpasses that of retail consumption scenarios. OSL Group holds Hong Kong Securities and Futures Commission Type 7 (Automated Trading Services) and Type 1 (Securities Deal) licenses, and is one of the few listed compliant exchanges globally with practical experience in both institutional custody and stablecoin settlement. Previously, in April 2026, it provided USDGO stablecoin settlement services worth approximately $130 million to institutional clients. This white paper provides an industry basis for the stablecoin policy framework of Hong Kong's financial regulators. Together with Visa's stablecoin settlement of $7 billion annualized scale, the launch of SoFi SoFiUSD, and the deployment of Western Union USDPT, it constitutes multiple evidence of the accelerated adoption of global corporate stablecoins.
Coinbase has obtained a MiFID license in the UK, which will allow it to offer derivatives and stock trading to UK users.
Odaily Odaily that Coinbase has announced it has obtained a UK Investment Services License (MiFID), enabling it to offer trading services for traditional financial products to UK users. Institutional and professional traders will be able to trade crypto, stock, and commodity perpetual contracts, while retail users will be able to trade stocks on the Coinbase platform for the first time. Coinbase stated that this license, together with its existing UK e-money license and crypto asset registration qualifications, forms part of its regulatory framework. It represents the largest expansion of the company's product capabilities since entering the UK market and will further advance its "Everything Exchange" strategy. In the future, the platform will also gradually support services such as stablecoin payments, savings, lending, and tokenized real-world assets (RWA).