![[ea9a93bd-e1c0-477a-a66a-8efed71bd42e.mp3]]

腾讯开源混元-A13B模型，个人及中小企业可低成本部署AI

简报：

• 腾讯正式开源了混元大模型家族新成员——混元-A13B模型，该模型采用专家混合（MoE）架构，总参数规模800亿，激活参数130亿。
• 混元-A13B模型在极端条件下仅需1张中低端GPU卡即可部署，降低了推理延迟和计算开销，便于个人开发者和中小企业接入前沿AI技术。
• 模型支持不同思考模式（快思考/慢思考），在数学、科学、逻辑推理任务中效果出色，并能完成复杂指令响应，助力智能体（Agent）应用开发。
• 用户可通过Github、HuggingFace等技术社区下载模型，API也在腾讯云官网同步上线。
• 为推动开源生态发展，腾讯还开源了两个全新数据集（ArtifactsBench和C3-Bench）支持代码与Agent场景模型评估。

相关链接：

• 腾讯混元-A13B开源新闻1
• 腾讯混元-A13B开源新闻2
• 模型体验入口
• 开源地址（Github）

Tencent Open-Sources Hunyuan-A13B Model, Enabling Low-Cost AI Deployment for Individuals and SMEs

Briefing:

• Tencent has officially open-sourced a new member of its Hunyuan large model family, the Hunyuan-A13B model. This model utilizes a Mixture-of-Experts (MoE) architecture with a total parameter size of 80 billion and 13 billion activated parameters. 🚀
• The Hunyuan-A13B model can be deployed with just one mid-to-low-end GPU card under extreme conditions, significantly reducing inference latency and computational overhead. This makes it easier for individual developers and small to medium-sized enterprises (SMEs) to access cutting-edge AI technology. 💡
• The model supports different thinking modes (fast thinking/slow thinking) and performs exceptionally well in mathematics, science, and logical reasoning tasks. It can also handle complex instruction responses, aiding in the development of AI Agent applications.
• Users can download the model through technical communities like GitHub and HuggingFace. The API is also simultaneously available on the Tencent Cloud official website.
• To foster the open-source ecosystem, Tencent has also open-sourced two new datasets (ArtifactsBench and C3-Bench) to support model evaluation in code and Agent scenarios. 🛠️

Related Links:

• Tencent Hunyuan-A13B Open-Source News 1
• Tencent Hunyuan-A13B Open-Source News 2
• Model Experience Portal
• Open-Source Address (Github)

OpenAI推出o3和o4-mini Deep Research API，助力自动化复杂研究与数据分析

简报：

• 2025年6月26日，OpenAI正式发布两款全新Deep Research API模型：o3-deep-research-2025-06-26与o4-mini-deep-research-2025-06-26，主攻高阶分析、深度信息合成及复杂自动化研究任务。
• o3模型着重复杂推理与精准分析，适用于金融分析、科学研究等高难度场景；o4-mini则兼顾效率与成本，擅长大规模快速查询。
• 两者均支持自动化网页搜索、文件分析、多步骤研究与代码执行等功能，通过API提供结构化、带引用的报告输出，并引入webhook机制提升异步处理效率。
• Deep Research API目前仅开放给ChatGPT生态内用户，以评估现实世界安全风险，并与谷歌Gemini Deep Research、DeepSeek R1等主流/开源竞品同场竞争。

相关链接：

• OpenAI发布Deep Research API新模型详情
• OpenAI深度研究API面市速报

OpenAI Launches o3 and o4-mini Deep Research APIs, Empowering Automated Complex Research & Data Analysis

Briefing:

• On June 26, 2025, OpenAI officially released two brand-new Deep Research API models: o3-deep-research-2025-06-26 and o4-mini-deep-research-2025-06-26. These models are primarily designed for advanced analysis, deep information synthesis, and complex automated research tasks. 🚀
• The o3 model focuses on complex reasoning and precise analysis, suitable for high-difficulty scenarios like financial analysis and scientific research; o4-mini balances efficiency and cost-effectiveness, excelling in large-scale rapid queries. 💡
• Both models support automated web searching, file analysis, multi-step research, and code execution. They provide structured, cited report outputs via API and introduce a webhook mechanism to enhance asynchronous processing efficiency.
• The Deep Research API is currently only available to users within the ChatGPT ecosystem to assess real-world safety risks, and it competes directly with mainstream/open-source rivals such as Google Gemini Deep Research and DeepSeek R1. 📈

Related Links:

• Details on OpenAI's New Deep Research API Models
• Flash Report: OpenAI Deep Research API Launched

谷歌开源轻量多模态模型Gemma3n，实现手机端AI性能突破

简报：

• 谷歌于6月27日正式发布并开源了Gemma3n多模态大模型，可在手机、平板、笔记本等设备本地运行，带来此前仅云端可用的多模态AI能力。
• Gemma3n包含E2B（5B参数，2GB内存）和E4B（8B参数，3GB内存）两个版本，架构创新使其仅占用传统2B/4B模型的资源，支持文本、音频、图片和视频等多模态输入及140种语言。
• 该模型采用MatFormer等多项创新架构，支持灵活组合、内存高效管理和KV Cache共享，提升多语言、数学、编码和推理处理效能。
• Gemma3n在LMArena基准上成为首个得分超1300的10B参数以下模型，并已在Hugging Face平台开放下载及文档，推动端侧AI技术应用新阶段。

相关链接：

• 谷歌Gemma3n官方Hugging Face页面
• Gemma3n官方文档
• IT之家相关报道

Mobile AI Performance Breakthrough with Google's Open-Sourced Lightweight Multimodal Model Gemma3n

Briefing:

• On June 27th, Google officially released and open-sourced the Gemma3n multimodal large model, enabling local execution on devices like smartphones, tablets, and laptops, bringing previously cloud-only multimodal AI capabilities directly to devices. 📱
• Gemma3n comes in two versions: E2B (5B parameters, 2GB memory) and E4B (8B parameters, 3GB memory). Its architectural innovations allow it to utilize resources comparable to traditional 2B/4B models, supporting multimodal inputs including text, audio, images, and video, as well as 140 languages. 🚀
• The model incorporates several innovative architectures, such as MatFormer, supporting flexible composition, memory-efficient management, and KV Cache sharing, thereby enhancing performance in multilingual processing, mathematics, coding, and reasoning. ✨
• Gemma3n is the first model with fewer than 10B parameters to score over 1300 on the LMArena benchmark. It is now available for download with documentation on the Hugging Face platform, ushering in a new phase for on-device AI technology applications.

Related Links:

• Google Gemma3n Official Hugging Face Page
• Gemma3n Official Documentation
• ITHome Related Report

蚂蚁开源代码图大模型登顶SWE-Bench，AI自动修复bug创开源新纪录

简报：

• 蚂蚁团队发布的开源大模型CGM，首次将代码仓库结构以图模态方式融入大模型，实现无需Agent即可自动修复bug，在SWE-bench Lite榜单以44%解决率排名开源模型第一，并整体性能媲美闭源方案。
• CGM以Qwen等开源基座为基础，通过图-语言多模态融合及Graph-RAG等创新，大幅提升模型对跨文件、跨模块软件工程任务的理解与自动化修复能力，显著优于此前开源水平。
• 该成果包括核心技术论文、代码、模型、数据集均已开源，突破了AI代码自动修复在透明性、私有化部署等层面的限制，为企业开发者提供了灵活、可控的新选择。

相关链接：

• 论文及开源资源汇总
• 开源代码仓库
• 新闻来源

Ant Group's Open-Source Code Graph Large Model Tops SWE-Bench, AI Automatically Fixes Bugs Setting New Open-Source Record 🚀

Brief:

• Ant Group's team has released the open-source large model CGM, which for the first time integrates code repository structures into a large model using a graph modality. This enables automatic bug fixing without the need for an Agent. CGM ranks first among open-source models on the SWE-bench Lite leaderboard with a 44% resolution rate, and its overall performance rivals closed-source solutions.
• Built upon open-source foundational models like Qwen, CGM significantly enhances the model's understanding and automated repair capabilities for cross-file and cross-module software engineering tasks through innovations such as graph-language multi-modal fusion and Graph-RAG, marking a substantial improvement over previous open-source efforts. ✨
• The achievements, including core technical papers, code, models, and datasets, have all been open-sourced. This breaks the limitations of AI code auto-repair in terms of transparency and private deployment, offering enterprise developers a flexible and controllable new choice. 🌟

Related Links:

• Paper and Open-Source Resources Summary
• Open-Source Code Repository
• News Source

DeepSeek R2模型因英伟达H20芯片短缺及出口限制开发受阻

简报：

• 受美国对华芯片出口限制影响，英伟达H20芯片在中国供应短缺，导致DeepSeek-R2 AI模型的开发出现延迟。
• DeepSeek团队原本依赖大批H20芯片进行R1与R2模型训练，目前因硬件受限，研发进度明显放缓。
• DeepSeek CEO梁文锋对R2模型现有能力仍不满意，内部正持续努力提升性能，尚未确定R2模型上线时间。
• 除新一代模型开发受到影响外，现有DeepSeek-R1模型的实际部署也因H20芯片短缺面临挑战。

相关链接：

• AI disruptor DeepSeek's next-gen model delayed by Nvidia H20 restrictions—short supply of accelerators hinders development
• DeepSeek's Next R2 Reasoning AI Model Stalls as a Casualty of US Chip War
• Report: DeepSeek's newest model delayed due to GPU export restrictions

DeepSeek R2 Model Development Hindered by Nvidia H20 Chip Shortage and Export Restrictions

Brief:

• The development of DeepSeek-R2 AI model faces delays due to a shortage of Nvidia H20 chips in China, impacted by US chip export restrictions. 📉
• The DeepSeek team, which originally relied on a large number of H20 chips for R1 and R2 model training, is now experiencing a significant slowdown in R&D progress due to hardware limitations. 🐌
• DeepSeek CEO Liang Wenfeng remains dissatisfied with the current capabilities of the R2 model, and the team is continuously working internally to improve its performance. The launch date for the R2 model is yet to be determined. 🛠️
• In addition to the impact on next-generation model development, the actual deployment of the existing DeepSeek-R1 model is also facing challenges due to the H20 chip shortage.

Related Links:

• AI disruptor DeepSeek's next-gen model delayed by Nvidia H20 restrictions—short supply of accelerators hinders development
• DeepSeek's Next R2 Reasoning AI Model Stalls as a Casualty of US Chip War
• Report: DeepSeek's newest model delayed due to GPU export restrictions