![[28fcfbfa-cda2-4d96-8437-aae5ebc901f9.mp3]]

MiniMax-M1发布：百万级上下文推理能力与极致性价比刷新开源AI新纪录

简报：

• 中国AI公司MiniMax正式开源其最新大模型MiniMax-M1，具备创纪录的100万token上下文窗口和8万token输出能力，是目前开源模型中上下文推理最长的之一。
• MiniMax-M1采用混合专家模型（MoE）架构与高效Lightning Attention机制，能够在复杂的长文档分析、代码生成、工具调用等场景中表现出色，性能媲美主流商业模型如OpenAI o3和Claude4Opus。
• 该模型以仅53.47万美元的训练成本和512张H800 GPU硬件支持，在三周内高效完成开发，相较于同类产品显著降低成本，实现业内领先的性价比。
• MiniMax-M1总参数4560亿，通过MoE架构带来高效计算，基于Apache 2.0协议已上架Hugging Face，全球开发者可免费下载和使用。
• 在软件工程、长上下文推理和多项综合测试中，MiniMax-M1表现优异，推动开源AI生态进一步发展。

相关链接：

• MiniMax-M1发布超长上下文推理能力，AI新王者仅53万美元打造
• MiniMax-M1开源：挑战国内外顶级AI模型
• MiniMax-M1技术细节与基准测试分析
• MiniMax公布超高性价比AI模型开发内幕
• 开源AI生态迎来欢呼，MiniMax-M1正式公开

MiniMax-M1 Launched: Million-Token Context Reasoning & Extreme Cost-Performance Redefine Open-Source AI Records

Brief:

• Chinese AI company MiniMax has officially open-sourced its latest large model, MiniMax-M1, featuring a record-breaking 1 million token context window and 80,000 token output capacity, making it one of the longest context reasoning open-source models currently available. 🤯
• MiniMax-M1 utilizes a Mixture-of-Experts (MoE) architecture and an efficient Lightning Attention mechanism, excelling in complex scenarios like long document analysis, code generation, and tool calling. Its performance rivals mainstream commercial models such as OpenAI o3 and Claude4Opus.
• Developed efficiently within three weeks, the model boasts an industry-leading cost-performance ratio, with a training cost of just $534,700 and support from 512 H800 GPUs, significantly reducing costs compared to similar products. 🚀
• MiniMax-M1, with a total of 456 billion parameters, achieves efficient computation through its MoE architecture. It is now available on Hugging Face under the Apache 2.0 license, allowing global developers to download and use it for free.
• MiniMax-M1 has demonstrated excellent performance in software engineering, long-context reasoning, and various comprehensive tests, further advancing the open-source AI ecosystem. ✨

Related Links:

• MiniMax-M1 Releases Ultra-Long Context Reasoning, New AI King Built for Only $530K
• MiniMax-M1 Open-Sourced: Challenging Top Domestic and International AI Models
• MiniMax-M1 Technical Details and Benchmark Analysis
• MiniMax Reveals Inside Story of Developing Ultra-Cost-Effective AI Models
• Open-Source AI Ecosystem Cheers as MiniMax-M1 is Officially Released

五角大楼与OpenAI达成2亿美元合作推动AI应对安全挑战

简报：

• 美国国防部已与OpenAI签订价值2亿美元的合同，目标是开发前沿人工智能原型，为作战和企业领域的关键国家安全挑战提供技术支持。
• 该项目主要在华盛顿及周边地区展开，预计于2026年7月完成。
• 合同背景为美国政府出台指导意见，虽明确将国防和国家安全系统排除在联邦AI指导意见之外，为此次合作提供政策支持。

相关链接：

• IT之家报道
• AIBase新闻1
• AIBase新闻2
• AIBase新闻3

Pentagon and OpenAI Ink $200 Million Deal to Advance AI for Security Challenges

Brief:

• The U.S. Department of Defense (DoD) has signed a $200 million contract with OpenAI, aimed at developing cutting-edge artificial intelligence prototypes to support critical national security challenges in both operational and enterprise domains. 🚀
• The project will primarily take place in Washington D.C. and its surrounding areas and is expected to be completed by July 2026. 🗓️
• This partnership is enabled by U.S. government guidance that explicitly excludes defense and national security systems from federal AI guidelines, thereby providing policy support for this collaboration. 🛡️

Related Links:

• IT Home Report
• AIBase News 1
• AIBase News 2
• AIBase News 3

Meta最新AI模型Llama 3.1在文本记忆测试中大幅提升表现

简报：

• 来自斯坦福大学等机构的最新研究显示，Meta于2024年发布的Llama 3.1（170B参数）模型在文本记忆能力上有显著提升，在回忆《哈利波特》第一部内容时，能够准确再现42%的内容。
• 相比之下，2023年的Llama 1（65B参数）仅能回忆起4.4%的同类内容，显示新版本模型在处理和记住经典书籍文本方面能力大幅提升。
• 该研究采用Books3数据库对多本受版权保护书籍进行测试，发现Llama 3.1在热门书籍的记忆能力普遍高于其他模型，此发现引发了关于AI与版权法规之间风险的新讨论。

相关链接：

• Meta's Llama 3.1 Memorized 42% of the first Harry Potter book
• AI Training and Copyright: Meta's Llama Model Literally Memorized 42% of Harry Potter, Study Finds
• Meta最新AI模型Llama 3.1能回忆哈利波特42%内容，研究显示大模型文本记忆能力显著增强

Meta's Latest AI Model Llama 3.1 Shows Significant Improvement in Text Memorization Tests

Brief:

• A new study from institutions including Stanford University reveals that Meta's Llama 3.1 (170B parameters) model, released in 2024, has shown a significant improvement in text memorization capabilities, accurately reproducing 42% of the content from the first Harry Potter book. 📚
• In contrast, the 2023 Llama 1 (65B parameters) model could only recall 4.4% of similar content, indicating a substantial enhancement in the new version's ability to process and remember classic book texts. 🚀
• The study used the Books3 database to test multiple copyrighted books and found that Llama 3.1 generally exhibited superior memorization of popular books compared to other models. This discovery has sparked new discussions regarding the risks between AI and copyright regulations. ⚖️

Related Links:

• Meta's Llama 3.1 Memorized 42% of the first Harry Potter book
• AI Training and Copyright: Meta's Llama Model Literally Memorized 42% of Harry Potter, Study Finds
• Meta最新AI模型Llama 3.1能回忆哈利波特42%内容，研究显示大模型文本记忆能力显著增强

阿里巴巴推出全面适配苹果MLX架构的Qwen3大模型

简报：

• 阿里巴巴集团宣布，其旗舰AI模型Qwen3已完成与苹果MLX架构的深度兼容优化，并一举开源32款MLX版本，覆盖从Mac到iPhone等全线设备。
• 新一代Qwen3模型包括4种精度（4bit、6bit、8bit、BF16），可为开发者提供灵活部署选择，标志着Apple Intelligence在中国正式上线前的关键推进。
• 据悉，苹果最终选择阿里巴巴作为在华AI合作伙伴，Qwen3系列可在苹果自有芯片平台实现高效训练和推理，加速全场景AI应用落地。

相关链接：

• 阿里巴巴发布适配苹果 MLX 架构的 Qwen3 升级版
• 阿里 Qwen3 全系列模型适配苹果 MLX 架构全开源
• Qwen3旗舰AI模型MLX版本开放下载，全面适配苹果芯片
• 苹果AI新进展，Qwen3 MLX支持多设备运行

Alibaba Launches Qwen3 Large Model Fully Compatible with Apple MLX Architecture

Briefing:

• Alibaba Group announced that its flagship AI model, Qwen3, has completed deep compatibility optimization with Apple's MLX architecture and open-sourced 32 MLX versions, covering a full range of devices from Mac to iPhone. 🚀
• The new generation Qwen3 model includes four precisions (4bit, 6bit, 8bit, BF16), offering developers flexible deployment options. This marks a critical step forward before the official launch of Apple Intelligence in China. ✨
• It is reported that Apple ultimately chose Alibaba as its AI partner in China. The Qwen3 series can achieve efficient training and inference on Apple's proprietary chip platforms, accelerating the implementation of AI applications across various scenarios. 📱

Related Links:

• Alibaba Releases Upgraded Qwen3 Version Compatible with Apple MLX Architecture
• Alibaba Qwen3 Full Series Models Fully Open-Sourced for Apple MLX Architecture
• Qwen3 Flagship AI Model MLX Version Available for Download, Fully Compatible with Apple Chips
• Apple AI New Progress: Qwen3 MLX Supports Multi-Device Operation

杭州举办国内首个人形机器人全产业链展，特斯拉等300家企业参展

简报：

• 2025杭州国际人形机器人与机器人技术展览会将于6月20日至22日在杭州大会展中心8号馆举行，展出面积达1.5万平方米。
• 展会吸引特斯拉、宇树等行业顶尖企业及阿里云、云深处、清宝等近300家知名企业参展，集中展示仿生驱动、具身智能、脑机接口等前沿技术及机器人上下游产业链。
• 展品涵盖机器人关键零部件、整机制造、系统集成等多个环节，期间还将举办国际人形机器人创新大会，探讨人形机器人关键技术和产业化进程。

相关链接：

• IT之家报道
• 36氪报道

Hangzhou to Host China's First Full Industry Chain Humanoid Robot Exhibition, Featuring Tesla and 300 Companies 🤖

Brief:

• The 2025 Hangzhou International Humanoid Robot and Robotics Technology Exhibition will take place from June 20th to 22nd at Hall 8 of the Hangzhou International Expo Center, covering an exhibition area of 15,000 square meters.
• The event has attracted industry-leading companies like Tesla and Unitree, along with nearly 300 well-known enterprises including Alibaba Cloud, Yunshenchu, and Qingbao. It will collectively showcase cutting-edge technologies such as bionic drives, embodied intelligence, brain-computer interfaces, and the entire upstream and downstream industrial chain of robotics. ✨
• Exhibits will cover various segments including key robot components, complete machine manufacturing, and system integration. An International Humanoid Robot Innovation Conference will also be held during the exhibition to discuss key humanoid robot technologies and their industrialization process. 🚀

Related Links:

• IT Home Report
• 36氪 Report