![[a905e088-6575-4cd9-98a7-dca7258114c6.mp3]]

阿里发布千问3向量模型，性能超谷歌等国际主流模型

简报：

• 2025年6月6日，阿里巴巴正式开源了基于千问3底座的Qwen3-Embedding向量模型系列，重点优化文本表征、检索和排序等任务。
• 新版本在文本检索、聚类和分类等核心任务上的性能较前代提升超40%，在MTEB等专业榜单上超越了谷歌的Gemini Embedding、OpenAI的text-embedding-3-large与微软同类产品，达到SOTA水平。
• 千问3向量模型支持超100种人类和编程语言，具备卓越的多语言和跨语言能力，可自定义指令或向量以优化特定场景，极大提升了开发者灵活性和效率。
• 该模型已在魔搭社区、Hugging Face、GitHub等平台开源，并可通过阿里云百炼以API方式调用，自4月底以来多次登顶国际榜单。

相关链接：

• 阿里全面开源千问3 Embedding与Reranker：性能超越Gemini、OpenAI等国际同类模型

• 阿里云开源千问3大模型：Embedding/Reranker能力登顶国际榜单

• 千问 Qwen3-Embedding：超越OpenAI和谷歌的文本向量新锐

Alibaba Releases Qwen3-Embedding Model, Outperforming Google and Other Major International Models

Briefing:

• On June 6, 2025, Alibaba officially open-sourced the Qwen3-Embedding model series, built upon the Qwen3 foundation model, with a focus on optimizing tasks such as text representation, retrieval, and ranking.
• The new version boasts over a 40% performance improvement over its predecessor in core tasks like text retrieval, clustering, and classification. It has surpassed Google's Gemini Embedding, OpenAI's text-embedding-3-large, and Microsoft's similar products on professional benchmarks like MTEB, achieving SOTA (State-of-the-Art) status. 🚀
• The Qwen3-Embedding model supports over 100 human and programming languages, demonstrating excellent multilingual and cross-lingual capabilities. It allows for custom instructions or embeddings to optimize specific scenarios, significantly boosting developer flexibility and efficiency. 🌍
• The model is now open-source on platforms such as ModelScope, Hugging Face, and GitHub, and can also be invoked via API through Alibaba Cloud Bailian. Since late April, it has repeatedly topped international leaderboards. ✨

Related Links:

• Alibaba Fully Open Sources Qwen3 Embedding and Reranker: Performance Surpassing Gemini, OpenAI, and Other International Counterparts
• Alibaba Cloud Open Sources Qwen3 Large Model: Embedding/Reranker Capabilities Top International Leaderboards
• Qwen3-Embedding: A New Text Vector Powerhouse Outperforming OpenAI and Google

谷歌Gemini2.5Pro预览版发布，编码能力领先主流AI模型

简报：

• 谷歌正式推出Gemini2.5Pro预览版，强调其在编码和推理能力方面超越DeepSeek R1、Grok 3 Beta、OpenAI等竞争对手，并在LMArena和WebDevArena等基准测试中获得高分。
• 新版本针对企业测试和新应用开发优化，响应速度更高、创意性和结构性改进明显，并已集成于多种新服务中。
• Gemini2.5Pro按照每百万tokens $1.25（输入）和$10（输出）计价，现可在Google AI Studio及Vertex AI平台使用。
• 业界和开发者对性能表现反应热烈，但实际提升仍需进一步长期验证。

相关链接：

• AIBase新闻1
• AIBase新闻2
• IT之家报道
• IT之家详细解读
• 功能介绍

Google Gemini 2.5 Pro Preview Released, Coding Capabilities Outpace Mainstream AI Models

Briefing:

• Google officially launched the preview version of Gemini 2.5 Pro, emphasizing its superior coding and reasoning capabilities over competitors like DeepSeek R1, Grok 3 Beta, and OpenAI. It achieved high scores in benchmarks such as LMArena and WebDevArena. 🚀
• The new version is optimized for enterprise testing and new application development, featuring faster response times, significant improvements in creativity and structural coherence, and has already been integrated into various new services. ✨
• Gemini 2.5 Pro is priced at $1.25 per million tokens for input and$10 for output, and is now available on Google AI Studio and Vertex AI platforms.
• The industry and developers have responded enthusiastically to its performance, though actual improvements require further long-term validation. 📈

Related Links:

• AIBase News 1
• AIBase News 2
• IT Home Report
• IT Home Detailed Analysis
• Feature Introduction

OpenAudio S1-Mini开源，轻量级AI语音模型支持多语言与丰富情感表达

简报：

• Fish Audio正式开源了全新文本转语音(TTS)模型OpenAudio S1-Mini，该模型拥有0.5B参数，支持14种语言及50多种情感和特殊语调，能够实现超自然、接近真人的AI语音合成。
• S1-Mini基于超过200万小时音频数据训练，从原有4B参数S1模型蒸馏而来，显著降低了计算门槛，便于在边缘设备和本地化应用部署。
• 该模型已在Hugging Face开放下载，限非商业用途，强化学习与人类反馈(RLHF)优化让其在自然度及多语言、情感表达等方面表现优异。
• OpenAudio S1-Mini的开源将促进AI语音技术的民主化，赋能小型团队和独立开发者，加快开源TTS技术在全球的推广与创新。

相关链接：

• AIbase新闻原文
• Hugging Face项目页

OpenAudio S1-Mini Open-Sourced: Lightweight AI Voice Model Supports Multilingual and Rich Emotional Expression

Briefing:

• Fish Audio has officially open-sourced its new Text-to-Speech (TTS) model, OpenAudio S1-Mini. This model features 0.5 billion parameters, supports 14 languages and over 50 emotions and special intonations, enabling ultra-natural, near human-like AI voice synthesis. 🗣️
• The S1-Mini was trained on over 2 million hours of audio data and distilled from the original 4 billion parameter S1 model, significantly lowering computational barriers and making it easier to deploy on edge devices and in localized applications. 🚀
• The model is now available for download on Hugging Face for non-commercial use. Its optimization through Reinforcement Learning with Human Feedback (RLHF) ensures outstanding performance in naturalness, multilingual capabilities, and emotional expression.
• The open-sourcing of OpenAudio S1-Mini is set to democratize AI voice technology, empowering small teams and independent developers, and accelerating the global adoption and innovation of open-source TTS technology. 🌐

Related Links:

• AIbase News Article
• Hugging Face Project Page

博世与阿里云共建智能座舱增强AI交互体验

简报：

• 德国博世携手阿里云，将通义大模型应用于智能座舱，提升汽车智能感知和3D数字人交互能力。
• 新方案借助AI技术，可自动感知座舱环境，智能调整温度、音乐和灯光，实现更自然的人车沟通体验。
• 采用云计算与端侧混合架构，支持语音、手势、表情等多模态控制，增强数据处理效率与安全。
• 智能座舱方案还集成了知识问答、同声传译等功能，为旅途提供实时服务和多语言交流支持。
• 此举标志着汽车智能化迈向以用户体验为中心的个性化移动空间。

相关链接：

• aibase新闻1

Bosch and Alibaba Cloud Collaborate to Build Smart Cockpit, Enhancing AI Interaction Experience 🚀

Briefing:

• Germany's Bosch partners with Alibaba Cloud to integrate the Tongyi Large Model into smart cockpits, enhancing intelligent perception and 3D digital human interaction capabilities in vehicles. 🚗
• The new solution leverages AI technology to automatically sense the cabin environment, intelligently adjust temperature, music, and lighting, achieving a more natural human-vehicle communication experience.
• It adopts a hybrid architecture of cloud computing and edge-side processing, supporting multi-modal control via voice, gestures, and facial expressions, enhancing data processing efficiency and security.
• The smart cockpit solution also integrates features like knowledge Q&A and simultaneous interpretation, providing real-time services and multi-language communication support for journeys. 🌍
• This initiative marks a significant step in automotive intelligence towards user-experience-centric personalized mobile spaces. ✨

Related Links:

• aibase News 1

面壁智能推出高效多场景端侧大模型MiniCPM 4.0

简报：

• 6月6日，面壁智能发布端侧大模型MiniCPM 4.0，包含8B和0.5B两个参数规模版本；
• MiniCPM 4.0-8B采用高效双频换挡机制，可自动切换稀疏和稠密注意力，优化长短文本场景性能；
• 0.5B版本被官方称为“轻巧灵动的最强小小钢炮”；
• 新模型支持vLLM、SGLang、LlamaFactory等主流框架部署，并搭载自研CPM.cu推理框架，实现极端场景下最高220倍、常规5倍推理提速和90%模型瘦身。

相关链接：

• ITHome报道

FaceWall AI Unveils Efficient Multi-Scenario Edge AI Model MiniCPM 4.0

Briefing:

• On June 6th, FaceWall AI released its edge AI model MiniCPM 4.0, available in two parameter scale versions: 8B and 0.5B.
• MiniCPM 4.0-8B incorporates an efficient dual-frequency shifting mechanism, allowing automatic switching between sparse and dense attention, optimizing performance for both long and short text scenarios. 🚀
• The 0.5B version is officially dubbed the "nimble and most powerful little steel cannon." 💪
• The new model supports deployment on mainstream frameworks like vLLM, SGLang, and LlamaFactory, and is equipped with FaceWall AI's self-developed CPM.cu inference framework, achieving up to 220x inference speedup in extreme scenarios, 5x in regular scenarios, and a 90% model size reduction. ✨

Related Links:

• ITHome Report