Scroller Briefings - HUAWEI Invests 1 Billion to Launch 'Tiangong Plan,' Accelerating HarmonyOS AI Ecosystem Development 🚀
编辑
目录
单词学习：![[8b47dacc-8bd2-499c-8052-5d992b2b4971.mp3]]
华为投资10亿启动天工计划，加速鸿蒙AI生态发展
简报：
- 在2025年华为全联接大会上，华为宣布正式启动“天工计划”。
- 该计划将投入10亿元人民币的资金与资源，旨在全面支持鸿蒙AI生态的创新。
- 华为将通过此计划支持开发者构建更多的AI元服务、意图框架和智能体，目标孵化超1万个AI原生元服务和5千多个智能体。
- 启动该计划时，华为同时宣布其HarmonyOS 5终端设备数量已突破1700万台。
相关链接：
HUAWEI Invests 1 Billion to Launch 'Tiangong Plan,' Accelerating HarmonyOS AI Ecosystem Development 🚀
Brief:
- At the 2025 HUAWEI CONNECT conference, Huawei officially announced the launch of the "Tiangong Plan."
- The plan will invest 1 billion RMB 💰 in funding and resources, aiming to fully support innovation within the HarmonyOS AI ecosystem.
- Through this plan, Huawei will support developers in building more AI meta-services, intent frameworks, and intelligent agents, targeting the incubation of over 10,000 AI-native meta-services and more than 5,000 intelligent agents. ✨
- Coinciding with the plan's launch, Huawei also announced that the number of HarmonyOS 5 terminal devices has exceeded 17 million units.
Related Links:
华为与浙大推出DeepSeek-R1-Safe模型，兼顾AI安全与性能
简报：
- 华为与浙江大学于华为全联接大会2025上，联合发布了基于昇腾千卡算力平台训练的DeepSeek-R1-Safe基础大模型。
- 该模型旨在解决AI的“安全与性能”难题，测试显示其对14个维度的有害问题防御成功率近100%，对多种越狱模式的防御成功率超过40%。
- 在提升安全性的同时，模型在MMLU、GSM8K等通用能力基准测试中的性能损耗控制在1%以内。
- 该模型目前已在Gitee、GitHub等多个社区全面开源。
相关链接：
Huawei and ZJU Launch DeepSeek-R1-Safe Model, Balancing AI Safety and Performance
Briefing:
- Huawei and Zhejiang University (ZJU) jointly unveiled the DeepSeek-R1-Safe foundational large model at Huawei Connect 2025. This model was trained on the Ascend thousands-card computing platform. 🚀
- The model aims to solve the "safety and performance" dilemma in AI. Tests indicate a nearly 100% defense success rate against 14 dimensions of harmful issues, and over 40% against various jailbreak modes.
- While enhancing safety, the model maintains excellent performance, with a less than 1% performance loss in general capability benchmarks such as MMLU and GSM8K. ✨
- The model is now fully open-sourced across multiple communities, including Gitee and GitHub. 🌐
Related Links:
- Huawei and ZJU Release DeepSeek-R1-Safe Foundational Large Model: Based on Ascend Thousands-Card Computing Platform, Solving AI "Safety and Performance" Dilemma
- Gitee Open Source Link
- GitHub Open Source Link
阿里云发布开源视频动作生成模型Wan2.2-Animate
简报：
- 阿里云近日宣布，旗下通义万相新一代动作生成模型 Wan2.2-Animate 正式全面开源。
- 该模型是Animate Anyone的全面升级版，在人物一致性和生成质量上有显著提升，支持动作模仿与角色扮演两种应用模式。
- 模型在视频生成质量、主体一致性等关键指标上超越了StableAnimator、LivePortrait等多个开源模型，在用户主观评测中表现甚至优于部分闭源模型。
- 用户可通过GitHub、HuggingFace和魔搭社区下载模型及代码，也可在阿里云百炼平台调用API或在通义万相官网直接体验。
相关链接：
Alibaba Cloud Releases Open-Source Video Action Generation Model Wan2.2-Animate
Brief:
- Alibaba Cloud recently announced that its new generation action generation model, Wan2.2-Animate, under Tongyi Wanxiang, has been officially and fully open-sourced. 🚀
- This model is a comprehensive upgrade of Animate Anyone, offering significant improvements in character consistency and generation quality. It supports two application modes: action imitation and role-playing.
- The model surpasses several open-source models like StableAnimator and LivePortrait in key metrics such as video generation quality and subject consistency. In user subjective evaluations, its performance even exceeds some closed-source models. ✨
- Users can download the model and code via GitHub, HuggingFace, and ModelScope communities. Alternatively, they can call the API on the Alibaba Cloud Bailian platform or experience it directly on the Tongyi Wanxiang official website. 💻
Related Links:
Mistral AI发布Magistral Small 1.2，支持视觉输入与128k上下文
简报：
- 法国AI公司Mistral AI正式发布其最新的开源推理模型Magistral Small 1.2，该模型拥有24B参数，并采用Apache 2.0开源许可。
- 新模型支持高达128k的上下文处理能力，并新增了视觉编码器，使其能够处理图像和文本的综合任务。
- 版本引入了一个创新的特殊token
[THINK]，用于包裹推理过程以提升模型表现。 - 模型内置多种推理模板，兼容vLLM、Transformers和llama.cpp等流行框架，并提供GGUF量化版本，方便开发者快速上手。
- 企业版Magistral Medium 1.2也同步升级，继续通过Le Chat平台和La Plateforme API提供服务。
相关链接：
Mistral AI Releases Magistral Small 1.2, Supporting Vision Input and 128k Context
Briefing:
- French AI company Mistral AI has officially launched its latest open-source inference model, Magistral Small 1.2. This model boasts 24B parameters and is released under the Apache 2.0 open-source license. 🚀
- The new model supports up to 128k context processing capability 🧠 and introduces a new vision encoder, enabling it to handle integrated image and text tasks.
- This version introduces an innovative special token
[THINK], designed to wrap the inference process to enhance model performance. - The model comes with various built-in inference templates, compatible with popular frameworks like vLLM, Transformers, and llama.cpp. It also provides GGUF quantized versions for developers to get started quickly.
- The enterprise version, Magistral Medium 1.2, has also been upgraded simultaneously, continuing to offer services via the Le Chat platform and La Plateforme API. ✨
Related Links:
- Mistral AI Launches Magistral Small 1.2 Open-Source Model
- Mistral AI Model Update Details
- AI Base Related Report
微软斥资40亿美元打造AI数据中心，宣称性能提升十倍
简报：
- 微软宣布投资40亿美元在美国威斯康星州建设一个新的AI数据中心。
- 该数据中心将配备数十万块英伟达Blackwell GB200芯片，用于支持复杂的AI训练和推理工作。
- 微软CEO萨蒂亚·纳德拉称，该数据中心的性能将是当前世界最快超级计算机的10倍。
- 为满足电力需求，微软计划建设一座太阳能发电场为数据中心提供绿色能源。
- 该项目中的首个数据中心预计将于2026年初投入使用。
相关链接：
Microsoft Invests $4 Billion in AI Data Center, Claims Tenfold Performance Boost
Briefing:
- Microsoft announced a $4 billion investment to build a new AI data center in Wisconsin, USA. 🚀
- The data center will be equipped with hundreds of thousands of NVIDIA Blackwell GB200 chips to support complex AI training and inference tasks. 🧠
- Microsoft CEO Satya Nadella stated that the data center's performance will be 10 times that of the world's fastest supercomputers today.
- To meet power demands, Microsoft plans to build a solar power farm to provide green energy for the data center. ☀️
- The first data center within this project is expected to be operational by early 2026.
Related Link:
英伟达斥资逾9亿美元收购Enfabrica团队，强化GPU互联技术
简报：
- 英伟达宣布以超过9亿美元的现金和股票，收购AI硬件初创公司Enfabrica的首席执行官（CEO）Rochan Sankar及其核心团队，并获得该公司的技术许可。
- Enfabrica成立于2019年，专注于开发能够高效连接超过10万块GPU的核心技术，以帮助构建大规模计算集群。
- 在此次收购前，英伟达曾于2023年参与Enfabrica的1.25亿美元B轮融资。
- 本次交易是科技行业通过收购顶尖AI人才以增强能力趋势的一部分，此前Meta和谷歌也有类似收购。
相关链接：
Nvidia Acquires Enfabrica Team for Over $900 Million to Strengthen GPU Interconnect Technology
Briefing:
- Nvidia announced the acquisition of AI hardware startup Enfabrica's CEO, Rochan Sankar, and his core team, along with a license to the company's technology, for over $900 million in cash and stock. 💰
- Founded in 2019, Enfabrica specializes in developing core technologies capable of efficiently connecting over 100,000 GPUs, aiming to build massive computing clusters. 🔗
- Prior to this acquisition, Nvidia had participated in Enfabrica's $125 million Series B funding round in 2023.
- This transaction is part of a broader trend in the tech industry of acquiring top AI talent to enhance capabilities, with Meta and Google having made similar acquisitions previously. 🚀
Related Links:
欧洲AI新突破，可提前十年预测千种疾病风险
简报：德国癌症研究中心等欧洲机构联合开发了一款名为Delphi-2M的生成式AI工具，该工具通过分析患者的病史、生活方式、年龄性别等数据，能够预测个人超过1000种疾病（如癌症、糖尿病、心脏病等）的潜在风险，并提前十年预测健康变化。
相关链接：
European AI Breakthrough: Predicts Risk of Thousands of Diseases Up to Ten Years in Advance
Brief: The German Cancer Research Center (DKFZ) and other European institutions have jointly developed a generative AI tool named Delphi-2M. This tool analyzes patient data including medical history, lifestyle, age, and gender, to predict an individual's potential risk for over 1000 diseases (such as cancer, diabetes, heart disease, etc.) and forecast health changes up to ten years in advance. 🩺💡
Related Links:
- German Cancer Research Center and Other Institutions Develop New AI Tool That Can Predict Risk for Over 1000 Diseases
- Nature Journal Research Paper 🔬
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