Scroller Briefings - MiniMax Open-Sources M1 Ultra-Long Context Large Model, Achieved for Just $530K with 1M Input Capability 🚀
编辑
目录
![[d042fefd-f3a7-4967-a2a0-8584066ae592.mp3]]
MiniMax开源超长上下文大模型M1，1M输入能力仅53万美元打造
简报：
- 中国AI公司MiniMax宣布开源其新一代大语言模型MiniMax-M1，具备业内领先的100万Token超长上下文输入和8万Token输出能力，远超同类开源模型，适用于复杂文档分析和多轮对话。
- MiniMax-M1采用MoE架构与Lightning Attention机制，强化代理任务表现，性能媲美OpenAI o3和Claude4Opus，训练成本仅为53.47万美元，显著低于行业主流大模型。
- 模型拥有456亿参数、支持多推理模式、基于Apache2.0协议开放，已上线Hugging Face和SiliconCloud平台，推动全球开源AI生态创新。
- 用户可通过SiliconCloud在线体验，首月新用户可获14元体验赠金。
相关链接：
MiniMax Open-Sources M1 Ultra-Long Context Large Model, Achieved for Just $530K with 1M Input Capability 🚀
Brief:
- Chinese AI company MiniMax has announced the open-sourcing of its new generation large language model, MiniMax-M1. It boasts an industry-leading 1 million Token ultra-long context input and 80,000 Token output capacity, significantly surpassing other open-source models. This makes it ideal for complex document analysis and multi-turn conversations.
- MiniMax-M1 employs a Mixture of Experts (MoE) architecture and a Lightning Attention mechanism to enhance agent task performance. Its capabilities are comparable to top models like OpenAI o3 and Claude4Opus, yet its training cost was a remarkably low $534,700, significantly undercutting mainstream large models in the industry. 💰
- The model features 45.6 billion parameters, supports multiple inference modes, and is openly available under the Apache 2.0 license. It has been launched on both Hugging Face and SiliconCloud platforms, aiming to foster innovation within the global open-source AI ecosystem. 💡
- Users can experience the model online through SiliconCloud, with new users receiving a 14 RMB trial credit for their first month.
Related Links:
月之暗面推出Kimi-Dev-72B，刷新AI编程基准测试纪录
简报：
- 月之暗面（Moonshot AI）发布了全新开源模型Kimi-Dev-72B，模型专注于软件工程任务，在AI编程基准测试SWE-bench Verified中取得了开源模型全球最高成绩。
- Kimi-Dev-72B以仅72亿的参数量取得60.4%的SWE-bench Verified测试得分，超越了此前671亿参数的DeepSeek-R1。
- 该模型通过大规模强化学习优化，能够自动修复真实代码库问题，结合BugFixer和TestWriter双重职能，提升代码修复与单元测试编写能力。
- Kimi-Dev-72B现已在Hugging Face（huggingface.co/moonshotai/Kimi-Dev-72B）和GitHub（github.com/MoonshotAI/Kimi-Dev）开源开放下载。
- 月之暗面计划未来继续优化该模型，并探索与主流开发工具的深度集成。
相关链接：
Moonshot AI Unveils Kimi-Dev-72B, Breaking AI Programming Benchmark Records
Briefing:
- Moonshot AI has released its new open-source model, Kimi-Dev-72B. This model is specifically designed for software engineering tasks and has achieved the highest global score among open-source models on the AI programming benchmark SWE-bench Verified. 🚀
- Kimi-Dev-72B scored 60.4% on the SWE-bench Verified test with only 7.2 billion parameters, surpassing the previous record held by DeepSeek-R1 (67.1 billion parameters).
- Optimized through large-scale reinforcement learning, the model can automatically fix issues in real codebases. It combines the dual functionalities of BugFixer and TestWriter, enhancing both code repair and unit test writing capabilities.
- Kimi-Dev-72B is now open-source and available for download on Hugging Face (huggingface.co/moonshotai/Kimi-Dev-72B) and GitHub (github.com/MoonshotAI/Kimi-Dev). 🌐
- Moonshot AI plans to continue optimizing the model in the future and explore deep integration with mainstream development tools. ✨
Related Links:
OpenAI升级ChatGPT搜索功能，提升智能响应和图像搜索能力
简报：
- OpenAI近期对ChatGPT的搜索功能进行重大升级，优化了搜索质量，使其智能响应更为精准。
- 新版本支持更好地理解和处理复杂对话上下文，并增强了指令跟踪能力，用户可通过更自然的方式提问，获得更准确的结果。
- 增加了图像搜索功能，用户可上传图片获取相关信息，丰富了搜索体验。
- 升级后的ChatGPT在生成回答时标注信息来源，提升用户核实信息的透明度。
- 对比传统搜索方式，此次升级意在更好地满足用户多样化的信息检索需求。
相关链接：
OpenAI Upgrades ChatGPT Search, Boosting Intelligent Responses and Image Search Capabilities
Brief:
- OpenAI recently made significant upgrades to ChatGPT's search functionality, optimizing search quality and making its intelligent responses more precise. 🚀
- The new version supports better understanding and processing of complex conversational contexts and enhances instruction-following capabilities, allowing users to ask questions more naturally and obtain more accurate results. ✨
- Image search functionality has been added, allowing users to upload pictures and obtain relevant information, thus enriching the search experience. 📸
- The upgraded ChatGPT annotates information sources when generating responses, enhancing transparency for users to verify information.
- Compared to traditional search methods, this upgrade aims to better meet users' diverse information retrieval needs.
Related Links:
OpenAI签署首个美国国防部AI合同，价值2亿美元
简报：
- OpenAI与美国国防部签署一项为期一年的2亿美元合同，目标是为国防部开发先进人工智能工具以应对国家安全挑战。
- 该项目将主要在华盛顿及周边地区推进，计划于2026年7月完成。
- 这是OpenAI首次与五角大楼合作，标志其扩大在国家安全领域的布局。
- 合同内容包括开发前沿AI原型，用于作战及企业关键应用。
- OpenAI目前年化收入已达100亿美元，本次合同虽金额一般但意义深远。
- 近年来，OpenAI、Anduril、Anthropic等多家公司在国防AI领域展开竞争和合作。
- 今年3月OpenAI新获软银等400亿美元融资，市场表现强劲。
相关链接：
OpenAI Signs First U.S. Department of Defense AI Contract, Valued at $200 Million
Briefing:
- OpenAI has signed a one-year, $200 million contract with the U.S. Department of Defense, aiming to develop advanced AI tools for the DoD to address national security challenges. 🚀
- The project will primarily be advanced in Washington D.C. and surrounding areas, with completion planned for July 2026.
- This marks OpenAI's first collaboration with the Pentagon, signaling its expansion into the national security sector.
- The contract involves developing cutting-edge AI prototypes for both operational and critical enterprise applications.
- While OpenAI's current annualized revenue has reached $10 billion, this contract, though moderate in value, carries profound significance. 💡
- In recent years, multiple companies including OpenAI, Anduril, and Anthropic have engaged in competition and collaboration within the defense AI sector.
- In March this year, OpenAI secured new funding of $40 billion from SoftBank and others, demonstrating strong market performance. 💪
Related Link:
谷歌发布Gemini 2.5 Flash-Lite模型，主打高效推理与低成本
简报：
- 谷歌于2025年6月18日正式推出全新轻量级AI模型Gemini 2.5 Flash-Lite，聚焦成本效益与推理速度提升。
- Flash-Lite版本在代码编写、科学计算、多模态分析等任务中，综合能力超越上一代2.0，尤其在翻译和分类等需要快速响应的场景表现更佳。
- 新模型支持灵活控制推理预算、外部工具连接，以及最长达100万token的超长上下文处理，有效助力生产级AI应用开发。
- 开发者可在Google AI Studio、Vertex AI等平台体验2.5 Flash和Pro的稳定版本，以及Flash-Lite的预览版；谷歌搜索等服务也已部署定制化版本以提升效率。
相关链接：
Google Launches Gemini 2.5 Flash-Lite Model, Emphasizing Efficient Inference and Low Cost
Brief:
- On June 18, 2025, Google officially launched its new lightweight AI model, Gemini 2.5 Flash-Lite, focusing on cost-effectiveness and enhanced inference speed. 🚀
- The Flash-Lite version surpasses its predecessor 2.0 in overall capabilities across tasks like code generation, scientific computing, and multimodal analysis, performing especially well in scenarios requiring quick responses such as translation and classification.
- The new model supports flexible control over inference budgets, external tool integration, and ultra-long context processing up to 1 million tokens, effectively aiding the development of production-grade AI applications. 💡
- Developers can experience the stable versions of Gemini 2.5 Flash and Pro, along with the preview version of Flash-Lite, on platforms like Google AI Studio and Vertex AI; customized versions have also been deployed in services such as Google Search to boost efficiency. ✨
Related Links:
前谷歌CEO支持的FutureHouse开源240亿参数化学推理模型ether0，准确率领先主流模型
简报：
- 由前谷歌CEO埃里克・施密特投资的FutureHouse初创公司，近日开源了一款名为ether0的化学任务推理模型，参数规模高达240亿，基于Mistral-Small-24B架构开发。
- ether0模型通过强化学习等多项技术，在涵盖18类化学任务的评测中表现优异，对比Claude、o1、ChemDFM和TxGemma等主流模型，在开放答案任务准确率居首，部分任务准确率甚至为同类两倍。
- 该模型无需额外领域预训练，对数据需求远小于传统模型，并能用自然语言推理和生成分子结构。
- ether0已为通用科学推理模型研究打下基础，重点在于类药分子设计，尽管目前仍处于原型阶段，但展现出强大潜力。
相关链接：
FutureHouse, Backed by Former Google CEO, Open-Sources 24 Billion Parameter Chemical Reasoning Model ether0, Leading Mainstream Models in Accuracy
Briefing:
- FutureHouse, a startup company invested in by former Google CEO Eric Schmidt, recently open-sourced
ether0, a chemical task reasoning model with a staggering 24 billion parameters, developed based on the Mistral-Small-24B architecture. 🧪 - The
ether0model, leveraging techniques like reinforcement learning, demonstrated outstanding performance in evaluations across 18 categories of chemical tasks. Compared to mainstream models such as Claude, o1, ChemDFM, and TxGemma, it achieved the highest accuracy in open-answer tasks, with some tasks even showing double the accuracy of its counterparts! ✨ - This model requires no additional domain-specific pre-training, has significantly lower data requirements compared to traditional models, and is capable of reasoning using natural language as well as generating molecular structures.
ether0has laid the groundwork for research into general scientific reasoning models, with a particular focus on drug-like molecule design. Although currently in its prototype phase, it demonstrates immense potential. 🚀
Related Link:
核电行业加速推动具身智能与人形机器人应用落地
简报：
- 银河证券最新研报显示，国内核电行业正在加速推动人形机器人和具身智能技术的产业化，已成为智能技术落地的重要应用场景。
- 行业内多家公司积极布局相关领域，包括景业智能深耕核工业应用、申昊科技与云深处科技联合开发核工业智能巡检机器人、上海电气启动“麒麟”具身智能训练场、亿嘉和推出RK100人形机器人探索核电巡检等。
- 核电领域以其危险、重复和繁重的作业特性，有望率先实现具身智能赋能及实际落地应用。
相关链接：
Nuclear Power Industry Accelerates the Application of Embodied AI and Humanoid Robots
Brief:
- Galaxy Securities' latest research report indicates that China's domestic nuclear power industry is rapidly advancing the industrialization of humanoid robots and embodied AI technologies, becoming a crucial application scenario for the deployment of intelligent technologies. 🤖
- Several companies within the industry are actively strategizing in related fields, including Kingyee Intelligent deeply cultivating nuclear industry applications, Shenhao Technology and Yunsenchu Technology jointly developing intelligent inspection robots for the nuclear industry, Shanghai Electric launching the "Kirin" embodied AI training ground, and Yijiahe introducing the RK100 humanoid robot for nuclear power inspection exploration. 💡
- Given its dangerous, repetitive, and heavy-duty operational characteristics, the nuclear power sector is expected to be among the first to achieve the empowerment and practical application of embodied AI. ☢️
Related Links:
如果对你有用的话，可以打赏哦
打赏
本文作者:topwind
本文链接:
版权声明:本博客所有文章除特别声明外，均采用 BY-NC-SA 许可协议。转载请注明出处！
目录