![[4943e000-7fb7-4387-b88b-4966c8a41e47.mp3]]

月之暗面Kimi K2模型再提速，最高每秒输出100 Tokens

简报：

• 8月22日，月之暗面（Moonshot AI）宣布其Kimi K2高速版模型（kimi-k2-turbo-preview）输出速度已提升，稳定速度达每秒60 Tokens，最高可达每秒100 Tokens。
• 该模型以MoE架构为基础，擅长代码生成和通用代理任务。
• 目前该模型提供五折特惠价格，活动将于2025年9月1日结束，届时将恢复原价。
• 优惠期间，模型每百万Tokens的输入价格为2元（缓存命中）或8元（缓存未命中），输出价格为32元。

相关链接：

• Kimi K2 高速版 AI 模型再提速，输出速度达每秒 100 Tokens
• kimi-k2-turbo-preview模型输出速度已提升至每秒 60 Tokens
• Moonshot AI 开放平台

Moonshot AI's Kimi K2 Model Accelerates Again, Reaching Up to 100 Tokens Per Second 🚀

Briefing:

• On August 22nd, Moonshot AI announced that the output speed of its Kimi K2 Turbo Preview model (kimi-k2-turbo-preview) has been enhanced, achieving a stable speed of 60 Tokens per second and a peak of up to 100 Tokens per second.
• Based on a Mixture-of-Experts (MoE) architecture, this model excels in code generation and general agent tasks.
• Currently, the model is offered at a 50% discount. This promotion will end on September 1, 2025, after which the original price will be reinstated. ⏳
• During the promotional period, the input price per million tokens is 2 RMB (cache hit) or 8 RMB (cache miss), while the output price is 32 RMB.

Related Links:

• Kimi K2 Turbo AI Model Accelerates Again, Output Speed Reaching 100 Tokens Per Second
• kimi-k2-turbo-preview Model Output Speed Enhanced to 60 Tokens Per Second
• Moonshot AI Open Platform ✨

苹果开源长视频AI模型，多项基准测试创纪录

简报：

• 苹果公司发布并开源了名为SlowFast-LLaVA-1.5的长视频多模态大语言模型，该模型在1B、3B、7B等多种参数规模下均刷新了LongVideoBench、MLVU等行业基准测试的最高纪录。

相关链接：

• 苹果新 AI 模型长视频理解夺冠，小至 1B 版本也领先对手

Apple Open-Sources Long Video AI Model, Sets Records in Multiple Benchmarks 🚀

Brief:

• Apple has released and open-sourced a long video multimodal large language model named SlowFast-LLaVA-1.5. This model, across various parameter scales including 1B, 3B, and 7B, has set new highest records in industry benchmarks such as LongVideoBench and MLVU. 🏆📹

Related Link:

• Apple's new AI model takes the lead in long video understanding, even the 1B version outperforms competitors

OpenAI深耕第二大用户市场，年内在印度开设首个办事处

简报：

• 人工智能公司OpenAI宣布，计划于年内在印度首都新德里设立其首个办事处，以深化与其第二大用户来源国的联系。
• OpenAI已在印度成立法人实体，并已开始在当地进行团队招聘。
• 此前，OpenAI已在印度市场推出了区域限定的低价订阅计划ChatGPT Go，月费为399印度卢比。
• OpenAI数据显示，过去一年ChatGPT在印度的周活跃用户数量增长了4倍，并且印度是其学生用户最多的国家。

相关链接：

• OpenAI 正式进驻印度，计划在新德里设立办公室
• OpenAI 年内开设首个印度办事处，深化与第二大用户来源国联系

OpenAI Deepens Roots in Second-Largest User Market, to Open First Indian Office This Year

Brief:

• AI company OpenAI announced its plan to establish its first office in New Delhi, the capital of India, this year. This strategic move aims to deepen its engagement with its second-largest user base country. 🇮🇳
• OpenAI has already established a legal entity in India and commenced local team recruitment. 🚀
• Previously, OpenAI launched the region-specific, low-cost subscription plan, ChatGPT Go, in the Indian market, priced at 399 Indian Rupees per month.
• According to OpenAI's data, ChatGPT's weekly active user count in India has quadrupled over the past year, and India is the country with its highest number of student users. 📈

Related Links:

• OpenAI Officially Enters India, Plans to Set Up Office in New Delhi
• OpenAI to Open First Indian Office This Year, Deepening Ties with Second-Largest User Source Country

特斯拉联手火山引擎与DeepSeek，升级中国市场语音助手

简报：

• 根据特斯拉中国官网更新的条款，其车机语音助手将接入字节跳动火山引擎提供的豆包大模型（Doubao）与人工智能公司DeepSeek的大模型。
• 升级后，豆包大模型将主要负责导航、空调调节等车辆核心语音命令，DeepSeek模型则专注于提供AI语音闲聊服务。
• 车主可通过物理按键或“嘿，Tesla”等唤醒词激活功能，全新上市的Model Y L车型将率先搭载此项升级。

相关链接：

• 特斯拉语音助手接入火山引擎与DeepSeek大模型 智能座舱体验全面升级
• 特斯拉接入豆包DeepSeek 可通过语音指令直接唤醒

Tesla Partners with Volcengine and DeepSeek to Upgrade Voice Assistant in China Market 🚗

Brief:

• According to updated terms on Tesla China's official website, its in-car voice assistant will integrate ByteDance's Volcengine Doubao large language model (LLM) and AI company DeepSeek's LLM.
• Following the upgrade, the Doubao LLM will primarily handle core vehicle voice commands such as navigation and air conditioning adjustments, while the DeepSeek model will focus on providing AI voice chat services. 💬
• Owners can activate the features via physical buttons or wake-up phrases like 'Hey, Tesla.' The newly launched Model Y L model will be the first to feature this upgrade. ✨

Related Links:

• 特斯拉语音助手接入火山引擎与DeepSeek大模型 智能座舱体验全面升级
• 特斯拉接入豆包DeepSeek 可通过语音指令直接唤醒

阿里通义发布Fun-ASR语音大模型，垂直行业识别准确率提升超15%

简报：

• 阿里通义与钉钉联合推出新一代语音识别大模型Fun-ASR，在家装、保险、畜牧等垂直行业，识别准确率提升超15%。
• 该模型采用自研语音算法与Qwen3微调技术，集成RAG检索增强方案，支持超1000个自定义热词导入，以优化特定场景的关键词识别。
• Fun-ASR引入强化学习技术以减少噪声干扰和识别错误，并对四川话、粤语等方言及复杂声学环境有较好适应性。
• 模型已集成至钉钉的会议字幕、智能纪要和语音助手等功能，并支持企业根据业务需求进行专属模型定制训练。

相关链接：

• 项目地址
• 相关报道1
• 相关报道2

Alibaba Tongyi Unveils Fun-ASR Large Speech Model, Boosting Recognition Accuracy by Over 15% in Vertical Industries

Brief:

• Alibaba Tongyi, in collaboration with DingTalk, has jointly launched the new generation Fun-ASR large speech recognition model, which boasts over a 15% improvement in recognition accuracy in vertical industries such as home decoration, insurance, and animal husbandry. 🚀
• The model utilizes self-developed speech algorithms and Qwen3 fine-tuning technology. It integrates a RAG (Retrieval-Augmented Generation) retrieval enhancement solution and supports the import of over 1,000 custom hotwords to optimize keyword recognition in specific scenarios. 💡
• Fun-ASR incorporates reinforcement learning technology to reduce noise interference and recognition errors, and demonstrates strong adaptability to dialects like Sichuanese and Cantonese, as well as complex acoustic environments.
• The model has been integrated into DingTalk's features such as meeting captions, smart summaries, and voice assistants, and also supports enterprises in training customized exclusive models based on their business needs. ✨

Related Links:

• Project Address
• Related Report 1
• Related Report 2

谷歌以0.47美元超低价向美政府提供AI，加剧科技巨头竞争

简报：

• 谷歌宣布将以每个机构0.47美元的特惠价格，向美国联邦政府提供其Gemini人工智能工具。
• 此价格低于竞争对手OpenAI和Anthropic此前提供的每机构1美元的折扣，加剧了政府AI服务市场的竞争。
• 该合作通过美国总务管理局（GSA）的“OneGov”采购策略实现，旨在简化联邦机构的采购流程。
• 谷歌的服务包中还包括视频和图像生成能力，但0.47美元的特惠价格仅在一年内有效，未来费用尚不明确。
• 有行业人士担忧，这种低价策略可能导致政府对单一供应商的依赖，从而造成未来更高的成本和风险。

相关链接：

• 谷歌以超低价向美国政府提供 AI 服务，促成行业竞争新格局
• 谷歌为美国政府提供 Gemini 人工智能工具，助力数字化转型

Google Offers AI to US Government at Ultra-Low Price of $0.47, Intensifying Tech Giant Competition 🚀

Brief:

• Google announced it will offer its Gemini AI tool to the US federal government at a special price of $0.47 per agency.
• This price undercuts competitors OpenAI and Anthropic, who previously offered discounts at $1 per agency, intensifying competition in the government AI services market.
• This partnership is facilitated through the US General Services Administration's (GSA) "OneGov" procurement strategy, designed to streamline purchasing for federal agencies.
• Google's service package also includes video and image generation capabilities, but the $0.47 special price is only valid for one year, with future costs remaining unclear.
• Industry insiders express concerns that this low-price strategy could lead to government reliance on a single vendor, potentially resulting in higher costs and risks in the future. 💸

Related Links:

• Google Offers AI Services to US Government at Ultra-Low Price, Creating a New Landscape of Industry Competition
• Google Provides Gemini AI Tools to US Government, Supporting Digital Transformation ✨