![[9f9faa59-f76d-411a-b7ba-ea9a573e7b48.mp3]]

Bing Image Creator免费引入GPT-4o模型

简报：微软旗下的Bing Image Creator现已免费提供由OpenAI先进的GPT-4o模型支持的AI图像生成服务。

相关链接：

• Bing Image Creator now offers OpenAI's GPT-4o image generation model for free
• 微软Bing Image Creator上线GPT-4o图像生成模型

Bing Image Creator Now Free with GPT-4o Model

Brief: Microsoft's Bing Image Creator now offers free AI image generation services powered by OpenAI's advanced GPT-4o model. 🎨✨

Related Links:

• Bing Image Creator now offers OpenAI's GPT-4o image generation model for free
• 微软Bing Image Creator上线GPT-4o图像生成模型

小红书开源自研多模态大模型，性能媲美顶尖闭源模型

简报：

• 小红书Hi Lab发布并开源了其首个自研多模态大模型dots.vlm1。
• 该模型基于12亿参数的NaViT视觉编码器和DeepSeek V3大语言模型，从零开始训练。
• 其核心亮点是自研的NaViT视觉编码器，支持动态分辨率，并善于处理表格、图表和文档等非典型结构化图片。
• 在MMMU、MathVision等国际评测基准上，dots.vlm1的综合表现已接近Gemini2.5Pro和Seed-VL1.5等领先的闭源模型。

相关链接：

• 小红书发布开源多模态大模型 dots.vlm1，以 NaViT 视觉编码器领跑行业

Xiaohongshu Open-Sources Self-Developed Multimodal Large Model, Performance Rivals Top Closed-Source Models

Briefing:

• Xiaohongshu Hi Lab has launched and open-sourced its first self-developed multimodal large model, dots.vlm1. 🚀
• The model is trained from scratch based on the 1.2 billion-parameter NaViT visual encoder and the DeepSeek V3 large language model.
• Its core highlight is the self-developed NaViT visual encoder, which supports dynamic resolution and excels at processing non-typical structured images like tables, charts, and documents. 📊
• On international evaluation benchmarks such as MMMU and MathVision, dots.vlm1's overall performance is already approaching leading closed-source models like Gemini2.5Pro and Seed-VL1.5. ✨

Related Link:

• Xiaohongshu Releases Open-Source Multimodal Large Model dots.vlm1, Leading the Industry with NaViT Visual Encoder

阿里推出Qwen3-4B，大模型迈向移动端

简报：

• 阿里巴巴发布了Qwen3-4B模型，该模型的一大亮点是支持在手机等终端设备上进行应用部署。

相关链接：

• 阿里发布Qwen3-4B模型，支持手机端应用

Alibaba Launches Qwen3-4B, Large Models March Towards Mobile 🚀

Brief:

• Alibaba has released the Qwen3-4B model. A key highlight of this model is its support for application deployment on terminal devices such as mobile phones. 📱✨

Related Link:

• Alibaba Releases Qwen3-4B Model, Supports Mobile Applications

腾讯开源WeKnora，提供基于大模型的多模态文档解决方案

简报：

• 腾讯正式开源了基于大语言模型（LLM）的文档理解与检索工具WeKnora。
• 该工具的核心能力是处理复杂的、包含文本、表格、图像的多模态文档（如PDF、Word），从中提取结构化内容并整合成统一的语义视图。
• WeKnora支持精准的智能问答和多轮对话交互，能深入理解用户意图，提升信息检索效率。
• 工具采用模块化架构，包含文档解析、向量化、检索引擎等组件，便于开发者根据企业知识库、科研、医疗、法律等不同场景进行定制与集成。

相关链接：

• 项目地址：Tencent/WeKnora

Tencent Open-Sources WeKnora, Offering Large Model-Based Multimodal Document Solutions

Briefing: 🚀

• Tencent has officially open-sourced WeKnora, a document understanding and retrieval tool based on Large Language Models (LLMs).
• Its core capability is processing complex multimodal documents (e.g., PDFs, Word files) containing text, tables, and images, extracting structured content, and integrating it into a unified semantic view. 🤯
• WeKnora supports precise intelligent Q&A and multi-turn conversational interaction, deeply understanding user intent and significantly improving information retrieval efficiency.
• The tool adopts a modular architecture, including components like document parsing, vectorization, and a retrieval engine, making it convenient for developers to customize and integrate it for various scenarios such as enterprise knowledge bases, scientific research, healthcare, and law. 🛠️

Related Link:

• Project Address: Tencent/WeKnora

AMD高通微软齐发力，推进OpenAI开源模型本地运行

简报：

• AMD与高通联合宣布，旗下硬件正式支持OpenAI新发布的gpt-oss系列开放模型。
• AMD锐龙AI Max+395处理器成为全球首款能运行gpt-oss-120b模型的消费级AI PC处理器。
• 高通表示，其骁龙平台在运行gpt-oss-20b模型时展现出色的思维链推理能力。
• 微软宣布，Windows 11将通过Windows AI Foundry平台，为用户提供gpt-oss-20b模型的本地运行支持。
• gpt-oss-20b模型可在配备16GB内存/显存的设备上运行，但OpenAI提示其“幻觉”比例较高，内部测试中约53%的回答存在事实错误。

相关链接：

• https://www.aibase.com/zh/news/20302
• https://www.aibase.com/zh/news/20312

AMD, Qualcomm, and Microsoft Join Forces to Advance Local Execution of OpenAI's Open-Source Models

Brief:

• AMD and Qualcomm jointly announced that their hardware officially supports OpenAI's newly released gpt-oss series of open models. 🚀
• The AMD Ryzen AI Max+ 395 processor is the world's first consumer AI PC processor capable of running the gpt-oss-120b model.
• Qualcomm stated that its Snapdragon platform demonstrates excellent chain-of-thought reasoning capabilities when running the gpt-oss-20b model.
• Microsoft announced that Windows 11 will provide users with local execution support for the gpt-oss-20b model through its Windows AI Foundry platform. 💻
• The gpt-oss-20b model can run on devices equipped with 16GB of RAM/VRAM, but OpenAI cautions about its high "hallucination" rate, with approximately 53% of answers containing factual errors in internal tests. ⚠️

Related Links:

• https://www.aibase.com/zh/news/20302
• https://www.aibase.com/zh/news/20312

网传GPT-5细节遭泄露，或含四大版本并强化智能体能力

简报：

• 一份疑似OpenAI旗舰模型GPT-5的详细说明信息在GitHub Models平台上被曝光，但OpenAI官方尚未对此回应。
• 泄露文件显示，GPT-5在推理能力和代码质量上有重大飞跃，并引入更强的“智能体能力”，能作为智能搭档协助用户完成多步骤复杂任务。
• 根据泄露信息，GPT-5预计将推出四个版本以适应不同场景：旗舰版gpt-5、轻量成本敏感版gpt-5-mini、低延迟版gpt-5-nano和企业级多模态对话版gpt-5-chat。
• 该泄露事件发生在OpenAI宣布将于北京时间8月8日（周五）凌晨1点举行直播活动之前，外界普遍猜测届时将正式发布GPT-5。

相关链接：

• 微软大漏勺！GPT-5信息提前泄露 将推出四种变体
• 重磅!OpenAI 旗舰模型 GPT-5详细信息疑似在 GitHub 上提前泄露
• GPT-5有望明天发布，将整合多项OpenAI的最新技术

GPT-5 Details Allegedly Leaked Online, May Feature Four Versions and Enhanced Agent Capabilities

Briefing:

• Detailed specifications for what appears to be OpenAI's flagship model, GPT-5, have reportedly been exposed on the GitHub Models platform. OpenAI has yet to officially respond to these claims. 🤫
• The leaked documents indicate that GPT-5 will feature significant leaps in reasoning capabilities and code quality, alongside the introduction of more robust "agent capabilities" that will allow it to act as an intelligent partner, assisting users with complex multi-step tasks. 💪
• According to the leaked information, GPT-5 is expected to launch in four versions to suit various scenarios: the flagship gpt-5, the lightweight cost-sensitive gpt-5-mini, the low-latency gpt-5-nano, and the enterprise-grade multimodal conversational gpt-5-chat.
• This leak occurred just before OpenAI announced a live stream event scheduled for 1 AM Beijing Time on Friday, August 8th, where it is widely speculated that GPT-5 will be officially unveiled. 🚀

Related Links:

• Microsoft's Big Leak! GPT-5 Information Leaked Early, Will Launch Four Variants
• Breaking! OpenAI's Flagship Model GPT-5 Detailed Information Allegedly Leaked Early on GitHub
• GPT-5 Expected to Launch Tomorrow, Integrating Multiple Latest OpenAI Technologies

AI语音新突破：FlowSpeech实现书面语向自然口语转换

简报：

• 人工智能语音合成工具FlowSpeech发布，其主要特点是能将书面文字转换为自然流畅的口语。
• 该技术通过上下文感知和多模态支持，深度理解文本语义，解决了传统TTS机械朗读、缺乏情感和语调变化的问题。
• FlowSpeech具备智能内容筛选功能，可自动识别并剔除广告、无意义字符串等干扰信息。
• 其应用场景包括播客制作、有声书、企业培训和教育材料，未来计划推出个性化声音定制服务。

相关链接：

• FlowSpeech：全球首个书面语转口语的 TTS

AI Voice Breakthrough: FlowSpeech Converts Written Language to Natural Spoken Speech

Briefing:

• The AI voice synthesis tool FlowSpeech has been released, with its main feature being the ability to convert written text into natural and fluid spoken language. 🗣️
• This technology deeply understands text semantics through context awareness and multimodal support, solving the problems of robotic reading, lack of emotion, and monotonous intonation in traditional TTS.
• FlowSpeech features intelligent content filtering, which automatically identifies and removes distracting information such as advertisements and meaningless character strings. ✨
• Its application scenarios include podcast production, audiobooks, corporate training, and educational materials, with future plans to introduce personalized voice customization services. 🎧

Related Links:

• FlowSpeech: The World's First Written-to-Spoken TTS