![[c40a7700-3081-4e96-a67d-121b664b53b4.mp3]]

Liquid AI开源高效LFM2模型，推动边缘AI性能和能效革新

简报：

• Liquid AI正式开源其下一代Liquid Foundation Models（LFM2）系列，该模型专为边缘设备设计，涵盖350M、700M和1.2B三种参数规模，采用创新的结构化自适应算子架构。
• LFM2推理速度比Qwen3快2倍，训练速度较前代提升3倍，在指令跟随和函数调用等关键任务中平均性能超越同类模型。
• 模型支持32K上下文长度、bfloat16精度、65K词汇表，并优化本地化隐私保护和低能源消耗，适用于智能手机、笔记本、汽车等边缘场景。
• LFM2模型权重已全部开源，开发者可在Hugging Face获取，并支持在Liquid Playground平台测试，进一步推动全球AI创新与普及。

相关链接：

• Liquid AI开源LFM2：边缘AI速度与效率双突破
• LFM2模型开源详情
• Liquid AI宣布LFM2模型架构创新
• 边缘计算与LFM2应用场景
• LFM2模型技术解读

Liquid AI Open-Sources Efficient LFM2 Model, Driving Edge AI Performance and Energy Efficiency Revolution

Briefing:

• Liquid AI has officially open-sourced its next-generation Liquid Foundation Models (LFM2) series, specifically designed for edge devices. This series includes models with 350M, 700M, and 1.2B parameters, utilizing an innovative Structured Adaptive Operator architecture. 🚀
• LFM2 boasts a 2x faster inference speed compared to Qwen3 and a 3x faster training speed than its predecessors. It also outperforms comparable models in average performance across critical tasks like instruction following and function calling. ✨
• The model supports a 32K context length, bfloat16 precision, and a 65K vocabulary. It is optimized for localized privacy protection and low energy consumption, making it ideal for edge scenarios such as smartphones, laptops, and automobiles. 💡
• All LFM2 model weights are now open-source and available for developers on Hugging Face. The models also support testing on the Liquid Playground platform, further promoting global AI innovation and adoption.

Related Links:

• Liquid AI Open-Sources LFM2: A Double Breakthrough in Edge AI Speed and Efficiency
• LFM2 Model Open-Source Details
• Liquid AI Announces LFM2 Model Architecture Innovation
• Edge Computing and LFM2 Application Scenarios
• LFM2 Model Technical Interpretation

亚马逊AWS发布Project Rainier大规模AI训练集群，算力提升五倍

简报：

• 亚马逊AWS于2024年6月24日介绍其大型分布式集群Project Rainier，该系统有望成为全球最强AI模型训练计算机之一。
• Project Rainier分布于美国多个数据中心，采用自研Trainium2 AI芯片，每个服务器配备16颗芯片，数以万计的UltraServer互联成超级集群。
• 集群内部通过NeuronLinks技术高速互联，数据中心之间则采用Elastic Fabric Adapter（EFA）技术。
• Anthropic将利用Project Rainier构建并部署Claude模型的下一代版本，AWS称其算力是当前Anthropic最大训练集群的五倍。

相关链接：

• IT之家新闻报道
• AIBase新闻1

Amazon AWS Unveils Project Rainier Large-Scale AI Training Cluster, Boosting Computing Power Fivefold

Briefing:

• On June 24, 2024, Amazon AWS introduced Project Rainier, a large-scale distributed cluster that is expected to become one of the most powerful AI model training computers globally. 🌍
• Project Rainier is distributed across multiple data centers in the U.S., utilizing proprietary Trainium2 AI chips 🧠. Each server is equipped with 16 chips, with tens of thousands of UltraServers interconnected to form a supercluster.
• Inside the cluster, high-speed interconnections are achieved via NeuronLinks technology, while Elastic Fabric Adapter (EFA) technology is used for connectivity between data centers.
• Anthropic will leverage Project Rainier to build and deploy the next generation of its Claude models. AWS states that its computing power is five times that of Anthropic's current largest training cluster. 🚀

Related Links:

• IT Home News Report
• AIBase News 1

Meta收购Play AI以增强AI语音与内容生成能力

简报：

• Meta宣布收购专注人工智能自然语音生成的初创企业Play AI，Play AI团队将于下周加入Meta。
• Play AI的语音创作平台与Meta在AI角色、Meta AI、可穿戴设备及音频内容创作等规划高度契合。
• 此举将增强Meta在智能语音、声音交互和内容创造领域的产品和创新能力，应对日益增长的AI语音技术需求。
• 收购财务细节未披露，Meta持续通过技术收购和顶尖AI人才招聘加快布局智能语音赛道。

相关链接：

• Meta收购AI语音初创公司Play AI
• Meta加速布局人工智能，收购Play AI团队

Meta Acquires Play AI to Enhance AI Voice and Content Generation Capabilities

Brief:

• Meta announced the acquisition of Play AI, a startup specializing in AI natural voice generation. The Play AI team is set to join Meta next week. 🚀
• Play AI's voice creation platform aligns perfectly with Meta's plans for AI characters, Meta AI, wearable devices, and audio content creation. ✨
• This acquisition will bolster Meta's product and innovation capabilities in intelligent voice, sound interaction, and content creation, addressing the growing demand for AI voice technology.
• Financial details of the acquisition were not disclosed. Meta continues to accelerate its efforts in the intelligent voice domain through technology acquisitions and recruitment of top AI talent. 💡

Related Links:

• Meta Acquires AI Voice Startup Play AI
• Meta Accelerates AI Deployment with Play AI Team Acquisition

Perplexity CEO高度认可月之暗面Kimi K2模型，考虑在其基础上深度合作

简报：

• 美国人工智能公司Perplexity的首席执行官Aravind Srinivas近日在社交媒体表示，对中国公司月之暗面最新开源的Kimi K2大模型表现高度赞赏，并公开透露未来可能会基于Kimi K2进行进一步训练和合作。
• Kimi K2采用混合专家（MoE）架构，总参数达1万亿，在代码生成和通用智能代理等任务及多个基准测试上均取得领先成绩，引发业内关注。
• Perplexity此前曾使用DeepSeek R1模型进行训练，此番对Kimi K2的积极评价，显示了其在AI模型合作和创新上的开放态度与期待。

相关链接：

• Perplexity CEO盛赞Kimi K2模型，称考虑后续训练
• 月之暗面发布万亿参数Kimi K2并开源
• Kimi K2模型亮点解析

Perplexity CEO Highly Praises Moonshot AI's Kimi K2 Model, Considering Deep Cooperation 🚀

Brief:

• Aravind Srinivas, CEO of U.S. AI company Perplexity, recently stated on social media that he highly praises the performance of Chinese company Moonshot AI's newly open-sourced Kimi K2 large model, and publicly revealed that future training and cooperation might be based on Kimi K2.
• Kimi K2 adopts a Mixture-of-Experts (MoE) architecture, with a total of 1 trillion parameters 🤖, achieving leading results in tasks such as code generation and general intelligent agency, as well as across multiple benchmarks, drawing significant industry attention.
• Perplexity previously used the DeepSeek R1 model for training; this positive evaluation of Kimi K2 demonstrates its open attitude and anticipation for AI model collaboration and innovation. ✨

Related Links:

• Perplexity CEO highly praises Kimi K2 model, considering future training
• Moonshot AI releases trillion-parameter Kimi K2 and open-sources it
• Highlights of Kimi K2 model explained

HuggingFace新推开源桌面机器人Reachy Mini，短时售额突破百万人民币

简报：

• AI平台HuggingFace近日发布开源小型桌面机器人Reachy Mini，主打教学、演示和机器人行为测试功能
• Reachy Mini高度28厘米，重1.5公斤，具备6自由度头部运动、全身旋转、广角摄像、多个麦克风与扬声器
• 其有线版（299美元起）和无线版（499美元起）受到市场欢迎，机器人开售五小时销售额已超13万欧元（约109万元人民币）
• 用户可通过HuggingFace社区下载、共享和扩展机器人动作，并与AI模型互动
• Reachy Mini核心采用Raspberry Pi5，体现HuggingFace以开源与低门槛理念推动机器人创新

相关链接：

• HuggingFace开源桌面机器人Reachy Mini五小时销售额超百万
• HuggingFace推出桌面机器人 Reachy Mini

HuggingFace Launches Open-Source Desktop Robot Reachy Mini, Sales Exceed One Million RMB in Short Order

Brief:

• AI platform HuggingFace recently released Reachy Mini, an open-source small desktop robot, primarily for teaching, demonstrations, and robot behavior testing. 🤖
• Reachy Mini stands 28 cm tall, weighs 1.5 kg, and features 6-DOF head movement, full-body rotation, wide-angle camera, multiple microphones, and speakers.
• Its wired version (starting at $299) and wireless version (starting at$499) have been well-received by the market, with sales exceeding 130,000 Euros (approximately 1.09 million RMB) within five hours of the robot's launch. 🚀
• Users can download, share, and expand robot actions through the HuggingFace community, and interact with AI models.
• Reachy Mini's core utilizes Raspberry Pi 5, reflecting HuggingFace's commitment to driving robot innovation through open-source and low-barrier principles. ✨

Related Links:

• HuggingFace开源桌面机器人Reachy Mini五小时销售额超百万
• HuggingFace推出桌面机器人 Reachy Mini