单词学习：![[2026042726838165.mp3]]

新闻播报：![[2026042735761145.mp3]]

OpenAI 开源本地隐私脱敏模型 Privacy Filter，支持 12.8 万 Token 与八类敏感信息识别

简报：

• OpenAI 推出前沿个人身份信息（PII）脱敏模型 Privacy Filter，已通过 Apache2.0 协议在 Hugging Face 和 GitHub 同步开源。
• 该模型具备深度语言理解能力，告别传统机械匹配，能根据上下文语境精准遮盖非结构化文本中的敏感信息并保留公开有用信息。
• 采用轻量化混合专家（MoE）架构，总参数15亿，每次推理仅激活约5000万参数，支持在笔记本和浏览器等边缘设备流畅运行。
• 具备 12.8 万 Token 的超长上下文窗口，在基准测试中取得 97.43% 的 F1 分数（召回率98.08%），确保长文本处理的连贯与准确。
• 支持识别八类核心敏感信息，涵盖基础身份（姓名/地址/电话/邮箱）、网络资产（URL）、金融账号、机密凭证（密码/API密钥）及日期。
• 该工具定位为云端大模型的本地“预过滤层”，实现数据不离设备的检测脱敏，防止隐私泄露；官方同时提醒医疗、金融等高敏感领域仍需人工审核。

相关链接：

• OpenAI 推出 Privacy Filter，支持12.8万超长上下文与8类隐私识别

OpenAI Open-Sources Local Privacy Redaction Model Privacy Filter, Supporting 128K Tokens and Eight Types of Sensitive Information 🔐

Brief:

• OpenAI has introduced Privacy Filter, a state-of-the-art personally identifiable information (PII) redaction model, and has simultaneously open-sourced it on Hugging Face and GitHub under the Apache 2.0 license.
• The model features deep language understanding capabilities, moving beyond traditional rule-based matching to accurately mask sensitive information in unstructured text based on contextual meaning while preserving publicly useful information.
• Built on a lightweight Mixture of Experts (MoE) architecture, it has 1.5 billion total parameters, with only about 50 million activated per inference, enabling smooth operation on edge devices such as laptops and browsers. 💻
• It supports an ultra-long context window of 128K tokens and achieved an F1 score of 97.43% in benchmark tests, with a recall rate of 98.08%, ensuring consistency and accuracy in long-text processing.
• It can identify eight core categories of sensitive information, including basic identity data (name/address/phone number/email), web assets (URLs), financial account information, confidential credentials (passwords/API keys), and dates.
• The tool is positioned as a local “pre-filtering layer” for cloud-based large models, enabling on-device detection and redaction so data never leaves the device, helping prevent privacy leakage. OpenAI also notes that highly sensitive sectors such as healthcare and finance still require human review. 🛡️

Related link:

• OpenAI Launches Privacy Filter with 128K Long Context and 8 Privacy Categories

DeepSeek全系API服务输入缓存命中价格大幅下调至一折

简报：

• DeepSeek官方于4月26日宣布，即日起其全系API服务的输入缓存命中价格降至原有价格的1/10。
• 针对Pro模型，用户在5月5日前还可额外叠加2.5折的限时优惠。
• 调整后，DeepSeek-V4-Flash模型的输入（缓存命中）价格降至0.02元，DeepSeek-V4-Pro模型的输入（缓存命中）价格降至0.025元起。

相关链接：

• DeepSeek API 输入缓存降价，降至原有价格 1/10

DeepSeek Slashes Input Cache Hit Pricing Across Its Entire API Lineup to One-Tenth 🚀

Brief:

• On April 26, DeepSeek officially announced that, effective immediately, the input cache hit pricing for its entire API product line has been reduced to one-tenth of the original price.
• For Pro models, users can also stack an additional limited-time 75% discount before May 5. 🎉
• Following the adjustment, the input (cache hit) price for the DeepSeek-V4-Flash model has been lowered to RMB 0.02, while the input (cache hit) price for the DeepSeek-V4-Pro model starts from RMB 0.025.

Related link:

• DeepSeek API input cache pricing cut to 1/10 of the original price

芯擎科技发布5nm车规级芯片“龙鹰二号”，支持大模型与舱驾融合

简报：

• 芯擎科技在2026年北京国际车展上正式发布自主研发的5纳米车规级AI座舱芯片“龙鹰二号”。
• 该芯片AI算力高达200TOPS，具备强大的性能冗余，能够原生支持7B以上参数的多模态大模型。
• 芯片采用极具前瞻性的“柔性架构”，能够覆盖从入门级到旗舰级车型的中央计算平台演进需求。
• “龙鹰二号”实现了“舱驾融合”的全场景覆盖，内部集成专用的车控处理单元与安全岛，可通过物理隔离安全处理座舱与驾驶辅助业务。
• 芯片计划于2027年第一季度正式启动车型适配工作，以加速智能汽车向中央计算时代的演进。

相关链接：

• 国产芯片再突破！5nm“龙鹰二号”正式发布，AI算力直达200TOPS

SiEngine Unveils 5nm Automotive-Grade Chip “Longying No. 2,” Supporting Large Models and Cockpit-Driving Integration 🚗🤖

Brief:

• At the 2026 Beijing International Auto Show, SiEngine officially unveiled its self-developed 5nm automotive-grade AI cockpit chip, “Longying No. 2.”
• The chip delivers up to 200 TOPS of AI computing power, offering substantial performance headroom and native support for multimodal large models with more than 7 billion parameters.
• Built on a highly forward-looking “flexible architecture,” the chip can meet the evolving central computing platform needs of vehicles ranging from entry-level to flagship models.
• “Longying No. 2” enables full-scenario coverage for “cockpit-driving integration,” with a dedicated vehicle control processing unit and safety island integrated on-chip, allowing cockpit and driving assistance functions to be handled securely through physical isolation.
• The chip is scheduled to officially begin vehicle integration and adaptation in the first quarter of 2027, accelerating the evolution of intelligent vehicles into the era of central computing. ⚙️

Related link:

• Another Breakthrough for Domestic Chips! 5nm “Longying No. 2” Officially Unveiled, Delivering 200 TOPS of AI Computing Power

腾讯开源具身多模态大模型 HY-Embodied-0.5-X，强化机器人智能交互能力

简报：

• 腾讯 Robotics X 实验室与混元团队共同发布并开源了专为具身任务优化的多模态大模型 HY-Embodied-0.5-X。
• 模型基于 HY-Embodied-0.5-MoT-2B 架构，核心强调机器人“看懂、想清、做到”的能力，在精细操作、空间推理、动作预测及风险判断上表现优越。
• 该系列包含两个主要版本：支持端侧实时响应的 MoT-2B 和支持更复杂任务处理的大参数规模 MoE-32B，主要面向家庭服务和桌面操作等实际应用场景。
• 数据与训练方面，结合了自采机器人第一视角操作数据与开源具身数据，引入思维链标注，并采用分阶段迭代的策略提升训练效率和稳定性。
• 该模型的推出进一步提升了机器人在空间理解、长程规划和具身交互方面的优势，推动机器人应用从“看懂”走向“干活”。

相关链接：

• 腾讯推出具身多模态大模型 HY-Embodied-0.5-X

Tencent Open-Sources the Embodied Multimodal Large Model HY-Embodied-0.5-X, Strengthening Robots’ Intelligent Interaction Capabilities 🤖✨

Brief:

• Tencent Robotics X Lab and the Hunyuan team have jointly released and open-sourced HY-Embodied-0.5-X, a multimodal large model specifically optimized for embodied tasks.
• Built on the HY-Embodied-0.5-MoT-2B architecture, the model emphasizes a robot’s ability to “see, think, and act,” delivering strong performance in fine manipulation, spatial reasoning, action prediction, and risk assessment.
• The series includes two main versions: the MoT-2B, which supports real-time on-device responses, and the larger-parameter MoE-32B, which is designed for handling more complex tasks. These models are primarily aimed at practical application scenarios such as household services and desktop operations.
• In terms of data and training, the model combines self-collected first-person robotic operation data with open-source embodied datasets, introduces chain-of-thought annotations, and adopts a phased iterative strategy to improve training efficiency and stability.
• The launch of this model further enhances robots’ strengths in spatial understanding, long-horizon planning, and embodied interaction, helping move robotic applications from simply “understanding” to actually “getting work done” 🚀

Related links:

• Tencent launches the embodied multimodal large model HY-Embodied-0.5-X

英伟达高管吴新宙展望自动驾驶：L4级量产即将成为现实

简报：

• 英伟达吴新宙表示，自动驾驶L4级的量产已经不再是梦想。

相关链接：

• 英伟达吴新宙：自动驾驶L4 级量产不再是梦

Nvidia Executive Wu Xinzhou Looks Ahead to Autonomous Driving: Mass Production of Level 4 Is About to Become Reality 🚗✨

Brief:

• Nvidia’s Wu Xinzhou said that mass production of Level 4 autonomous driving is no longer just a dream.

Related Links:

• Nvidia’s Wu Xinzhou: Mass Production of Level 4 Autonomous Driving Is No Longer a Dream

谷歌拟最高掷400亿美元加码Anthropic，提供百万颗TPU并助推其估值达3500亿

简报：

• 谷歌计划向AI初创公司Anthropic注入100亿美元资金，根据双方长期规划，未来投资总额或将攀升至400亿美元。
• 此次巨额交易预计将把Anthropic的估值推高至3500亿美元。此前在今年2月，该公司刚完成300亿美元融资，而亚马逊也曾投入50亿美元并保留200亿美元的追加投资计划。
• 双方深度捆绑底层硬件资源，谷歌将在未来五年内为Anthropic提供5吉瓦的算力支持以及100万颗自主研发的TPU芯片。
• 尽管双方达成深度合作，但谷歌高层对Anthropic在AI编程市场的迅速崛起仍抱有担忧，未来在特定细分领域可能演变为直接竞争。
• Anthropic目前正积极筹备资本市场的下一步动作，有消息称其正考虑在今年10月正式启动IPO进程。

相关链接：

• 谷歌拟加码投资 Anthropic，总金额或达400亿美元

Google Plans to Invest Up to $40 Billion More in Anthropic, Provide 1 Million TPUs, and Help Boost Its Valuation to$350 Billion 🚀

Brief:

• Google plans to inject dJyH6kDkfC5RlM16.answerText0 billion into AI startup Anthropic, and according to the long-term roadmap between the two parties, the total investment could eventually rise to $40 billion.
• This massive deal is expected to push Anthropic’s valuation up to $350 billion. Earlier this February, the company had just completed a$30 billion funding round, while Amazon had also invested $5 billion and retained plans for an additional$20 billion investment.
• The two companies are deeply aligned on underlying hardware resources, with Google set to provide Anthropic with 5 gigawatts of computing power support and 1 million self-developed TPU chips over the next five years.
• Despite the deep partnership, Google executives remain concerned about Anthropic’s rapid rise in the AI coding market, and the two could eventually become direct competitors in certain niche segments.
• Anthropic is currently actively preparing its next move in the capital markets, with reports suggesting that it is considering officially launching its IPO process in October this year 📈.

Related link:

• Google Plans to Increase Investment in Anthropic, with Total Amount Potentially Reaching $40 Billion