模型发布/更新
-
面壁智能发布首个具身智能成果 MiniCPM-Robot 系列模型,含 1.5B VLA 与 0.9B 跟踪模型
面壁智能联合 OpenBMB 发布并开源 MiniCPM-Robot 系列,包括通用 VLA 模型 MiniCPM-RobotManip(1.5B 参数)与移动跟踪模型 MiniCPM-RobotTrack(0.9B 参数)。
-
NVIDIA releases Cosmos 3 Edge: a 4B-parameter open-source world model for real-time reasoning and action generation in robots and edge AI
NVIDIA open-sourced Cosmos 3 Edge on Hugging Face, a 4-billion-parameter world model designed to help robots and…
-
通义实验室发布 Qwen-Audio-3.0-TTS 实时语音合成模型
通义实验室发布 Qwen-Audio-3.0-TTS,含 Flash(首包延迟约300ms)和 Plus 两个版本。Plus 版本在 Artificial Analysis 榜单夺冠,支持16种语言和20种中文方言,平均 WER/CER…
-
Shanghai Academy of AI for Science opens multimodal science foundation model "Shenzhen"
The Shanghai Academy of AI for Science has opened "Shenzhen", a multimodal science foundation model with about 11…
产品发布/更新
-
xAI launches Grok for Excel add-in for natural-language data analysis and formula writing
xAI released a Grok add-in for Microsoft Excel that lets users ask questions in natural language, write formulas from…
-
Xiaohongshu and Peking University open-source UltraEP: real-time load balancing for large-scale MoE training and inference
Xiaohongshu and Peking University proposed UltraEP, the first to bring real-time load balancing based on precise…
-
Cursor tests new AI agent cluster: planner plus executor, 80% SQL tests passed in 4 hours
Cursor tested a new AI agent cluster that splits tasks between a planner, using the strongest model, and executors,…
-
Ray 2.55 正式支持 Google Cloud TPU,通过 KubeRay 自动编排多主机切片
Ray 2.55 首次为 Google Cloud TPU 提供一等支持,开发者可通过 Ray 任务与 Actor API 在 TPU 上运行分布式 Python 负载。
-
Claude Code v2.1.216 发布:修复长会话卡顿与多项 Agent 行为问题
Claude Code v2.1.216 修复了长会话中消息归一化成本随轮次二次增长导致的数秒停顿问题,并修正了 OAuth token 过期后自动模式误判 HTTP 401、后台子智能体重启后恢复默认配置、以及工作树隔离子智能体重定向…
行业动态
-
District 9 director Neill Blomkamp releases Nightborne, first fully AI-generated short film
Neill Blomkamp released Nightborne, a 13-minute sci-fi horror short film created entirely frame by frame with the…
-
Hugging Face breached by autonomous AI agent, uses AI tools for forensic analysis in hours
Hugging Face disclosed that part of its production infrastructure was breached by an autonomous AI agent system. The…
-
Ollama raises $88M to accelerate open model ecosystem
Ollama announced it has raised $88 million in funding led by Benchmark, Theory Ventures and 8VC. The platform serves…
论文研究
-
Apple 提出 Length Value Model (LenVM):token 级长度建模框架
Apple 研究团队提出 LenVM,一种在每步解码时预测剩余生成长度的 token 级框架,将长度建模转化为无需标注的价值估计问题。在 LIFEBench 精确长度匹配任务上,LenVM 将 7B 模型的长度分数从 30.9 提升至 6…
-
LVSum 基准:评估多模态大模型的长视频时间感知摘要能力
Apple 推出 LVSum,一个含细粒度时间对齐的人工标注长视频摘要基准,包含 13 个领域的 72 个视频(平均时长 16 分钟),每个视频配有最多 10 条含时间引用的人类摘要。评估显示,转录文本对摘要质量的贡献远大于视觉帧,当前多…
技巧与观点
-
Build products without code: a step-by-step Vibe Coding tutorial from scratch
This article is aimed at zero-code users and provides a complete process for developing and launching a product from…
-
OpenAI finds new failure modes in long-running model safety and alignment practice, improves evaluation system
While internally using a long-running model that can operate autonomously for hours to weeks, OpenAI observed new…
-
China's AI nearly catches up with US as open-source Kimi K3 shakes the market
Chinese company Moonshot.AI released the Kimi K3 model, whose performance matches the best US models and which is…
-
开源权重模型逼近前沿,闭源仍领先
开源权重模型已多次达到与闭源前沿模型相当的水平,但闭源模型如 GPT-5.2 仍持续创造阶跃式突破。Kimi K3 为 2.8T 参数开源模型,Qwen 3.8 为 2.4T 参数模型。按 90/10 输入输出比例计算,开源前沿模型价格中…
-
乐天用 Claude Fable 5 构建可自主运行数小时的智能体
乐天AI业务总经理Yusuke Kaji测试Claude Fable 5后表示,该模型能自主运行更长时间,首次实现夜间无人值守完成复杂任务。与之前模型不同,Fable 5在运行中持续自我验证和纠错,避免早期错误假设导致整次运行失败。乐天已…