OpenWiki:为编码智能体生成并维护开源仓库文档
LangChain 推出 OpenWiki,一个为编码智能体自动生成和维护代码库文档的开源工具。它让智能体能直接检索所需的仓库上下文,无需将所有内容加载到单个指令文件中。
推荐理由:LangChain 开源的这个文档生成代理解决了编码代理上下文管理的一个痛点,对重度使用 agent 做开发的团队很实用。
LangChain 推出 OpenWiki,一个为编码智能体自动生成和维护代码库文档的开源工具。它让智能体能直接检索所需的仓库上下文,无需将所有内容加载到单个指令文件中。
推荐理由:LangChain 开源的这个文档生成代理解决了编码代理上下文管理的一个痛点,对重度使用 agent 做开发的团队很实用。
葡萄牙政府7月1日宣布开源大语言模型AMALIA正式发布,这是首个基于欧洲葡萄牙语开发的模型。由60余位研究人员历时18个月开发,先期投资550万欧元,使用了Deucalion、MareNostrum 5等算力基础设施。第一阶段利用约4万亿个葡语单词训练出9B规模模型,后续多模态升级后具备理解文本、图像、声音的能力。今年内还将新增22B版本并获得智能体能力,需追加150万欧元投资。
Lift 是一个 PDF 到结构化数据的提取工具,本教程构建了完整的受控评估工作流。通过 4-bit NF4 量化,约 10B 参数模型可在 16 GB GPU(如 T4/L4)上运行。教程生成含干扰项的合成多页研究论文,测试模型在真实文档布局中恢复标题、作者、数据集、指标、超参数、局限性和仓库链接的能力,并输出结构化 JSON,实现模式引导的字段级评估。
U.S. chip restrictions helped push China to build and spread open AI models. The authors tested this by looking at policy documents, open model releases, GitHub activity, research papers, company-linked papers, and U.S. patents. They found that after major U.S. export controls, Chinese developers increased activity around open LLM projects much more than U.S. developers did. ---- Link – arxiv. org/abs/2606.15999 Title: "U.S.Policies Unintentionally Accelerated China's Open AI Ecosystems"
Venice AI 获得由 Dragonfly 领投、Coinbase Ventures 等参投的 6500 万美元 A 轮融资,估值 10 亿美元。该平台提供 200 余个 AI 模型,注重隐私:用户输入加密且不存储数据,部分模型支持端到端加密(需付费订阅)。目前拥有 85 万独立访问用户、300 万活跃用户,日均 170 万次 API 调用,年化收入超 7000 万美元,已实现盈利。公司发行了 VVV 和 DIEM 两种加密代币,约 8% 用户使用加密支付。融资将用于购买 GPU 并自建数据中心。
瑞银分析师调研十余家企业IT负责人后发现,约六成企业已出台管控措施收紧AI开支。token调用成本是核心顾虑,Uber运营总监坦言AI投入回报微薄。DeepSeek等开源模型与中国本土大模型有望受益,尤其适合非代码类业务。短期内OpenAI、Anthropic等闭源厂商承压最大。谷歌推出Gemini 3.5 Flash,Anthropic发布Claude Sonnet 5以降低成本。瑞银称这一趋势为“良性调整阵痛”,无企业彻底叫停AI落地。
美团发布LongCat-2.0,总参数1.6万亿、每个token仅激活约480亿参数的MoE模型,从预训练到大规模部署全程运行在5万张国产算力芯片上,训练消耗超35万亿tokens,无回滚、无不可恢复loss突刺。Agent场景表现突出:在Terminal-Bench 2.1和SWE-bench Pro编程任务上追平Gemini 3.1 Pro,FORTE通用Agent任务与Claude Opus 4.6持平。最大输出128K,最高提供1M上下文,采用LSA稀疏注意力机制和N-gram Embedding优化长上下文与工具调用。已开放API接入,支持OpenAI Compatible和Anthropic API生态,可直接接入Claude Code和Codex等工作流。
NVIDIA 发布 Nemotron-Labs-TwoTower,基于冻结的自回归骨干 Nemotron-3-Nano-30B-A3B 的扩散语言模型。采用双塔架构:上下文塔冻结,降噪器塔训练,通过层对齐交叉注意力和状态播种协作。在 2×H100 上 BF16 评估,保留 98.7% 的 AR 基线质量,生成吞吐量提升 2.42 倍(γ=0.8,块大小 S=16)。降噪器在约 2.1T token 上训练,骨干使用 25T token 预训练。总参数约 60B,每 token 活跃参数约 3B/塔。支持扩散、模拟 AR 和 AR 三种解码模式。
推荐理由:NVIDIA这个TwoTower把扩散解码接在已有的AR骨干上,几乎无损质量却让吞吐翻倍,并且开源可商用,对批量文本生成的团队是实在的加速工具。
Godot 基金会正式修改贡献指南,禁止开发者提交 AI 直接生成的代码、AI 智能体发起的 Pull Request,以及沟通过程中使用 AI 翻译器生成的文本。理由是近几个月“AI slop”问题持续消耗维护者精力、打击审查者积极性。新规允许 AI 用于“机械性的小任务”,但必须主动披露使用情况。所有贡献必须由能对代码负责的人类提交,并在出现问题时承担修复责任。目前基金会表示将采取保守态度,未来政策可能随技术发展重新评估。
Google Research 发布 TabFM v1.0.0,一种面向表格数据的零样本基座模型,无需针对特定数据集训练、调参或特征工程即可进行分类和回归。TabFM 将整个数据集作为统一提示,通过上下文学习在单次前向传播中生成预测。其架构融合 TabPFN 风格的行/列交替注意力与 TabICL 风格的行压缩和上下文学习。模型在数亿个由结构因果模型动态生成的合成数据集上训练。在 TabArena 基准(38 个分类、13 个回归数据集,样本量 700 至 15 万)上,TabFM 及添加交叉特征和 SVD 特征、采用 32 路集成的 TabFM-Ensemble 均优于经过深度调优的监督算法。模型已开源在 Hugging Face 和 GitHub,Google BigQuery 即将通过 `AI.PREDICT` SQL 命令集成。
Opinion from a former Meta PM. And this is from Aravind Srinivas of Perplexity "China can build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem." https://x.com/rohanpaul_ai/status/2071780740220752220/video/1
Backstage and onstage with @Thom_Wolf and @swyx . I really enjoyed the fireside chat! Thanks for having me back at @aiDotEngineer! And always proud to be part of these conversations to share our work on sparse attention and native multimodality trained from the start, and why we believe open weights matter.
tufalabs 刚刚开源了他们获得第一名的 notebook。主推文表示,这有望解锁新一波更好的参赛作品。
.@tufalabs just open sourced their 1st place notebook 👀
美团于6月30日发布新一代万亿参数大模型LongCat-2.0并开源。总参数1.6T,平均激活约48B,原生支持1M超长上下文,在五万卡国产算力集群上完成全流程训练与推理。采用LSA稀疏注意力、零计算专家、ScMoE及MOPD多专家融合(Agent/Reasoning/Interaction三组专家)架构。评测中SWE-bench Pro获59.5,SWE-bench Multilingual获77.3。预览版已通过OpenRouter和longcat.ai开放,月调用量跻身OpenRouter全球前三。
推荐理由:国产算力上首个全流程自训的万亿开源模型,1M上下文和动态专家架构直指Agentic Coding场景,OpenRouter调用量已经冲到前三,不是Demo是生产力。
Factory 发布 Droid Shield 2.0,在确定性密钥扫描器两侧各加一个微调模型:risk 模型在扫描器未触发时按 recall 优先提示可能漏检的密钥,downgrade 模型对扫描命中做脱敏后判断是否误报。
Together AI 宣布完成 8 亿美元 C 轮融资,投资方包括 Aramco Ventures、NVIDIA、Vista Equity、General Catalyst 等,并获超过 500 MW 算力容量承诺。
前沿AI模型能力加速提升,美国实验室发布速度加快,但政府干预已限制访问Claude Fable和GPT-5.6。Epoch测试发现Opus 4.7自主运行14小时即可完成需2-17周人工的软件工程,token成本$251。中国开源模型落后前沿6-12个月,但性能也快速提升,在AA-Briefcase测试中呈独立指数曲线。使用方式正从聊天机器人转向智能体,OpenAI内部四分之一员工每周同时运行至少四个智能体。Claude Code用户数据显示,领域经验比职业属性更决定使用效果,专家正用智能体替代此前非专家的聊天机器人使用模式。
推荐理由:Ethan Mollick 用数据和亲身实验把话说得很直白,AI 不再是聊天工具而是能独自干上 9 小时活的 agent。做非技术岗的也在用,工作变成管理 AI,这个转变才刚开始,但已经让机构跟不上。
免费开源 AI 智能体 OpenClaw 现已推出 iOS 和 Android 应用。用户通过 OpenClaw Gateway 将手机与智能体及所需工具、技能连接,可在移动端运行智能体完成编程、餐食规划等任务。OpenClaw 今年早些时候因 MoltBook(号称完全由智能体运营的社交媒体站)走红,其创建者 Peter Steinberger 随后加入 OpenAI。尽管 MoltBook 后被证实部分由人类伪装智能体运行,但此事件推动了智能体生态扩张,如今 AI 智能体已嵌入更多场景,包括手机端。
Step 3.7 Flash 本月在 @OpenRouter 上跻身前十,路由了 4.29T tokens。 构建者正在通过真实的智能体运行、编码任务和长上下文工作流来推动它。 继续用难题来挑战。
Opinion from a former Meta PM. And this is from Aravind Srinivas of Perplexity "China can build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem." https://x.com/rohanpaul_ai/status/2071780740220752220/video/1
Deepseek推出DSpark推理框架,采用推测解码技术,由小模型生成候选答案、大模型批量验证,并一次生成多个token而非单个,使每用户响应速度提升60–85%。系统基于置信度动态调整验证深度,减少无效计算。DSpark与Deepseek-V4-Pro模型(与北京大学联合开发)已在HuggingFace和GitHub以MIT许可证开源。高效推理降低对高端芯片需求,有助于中国与欧盟在芯片受限下获取更多AI性能,短期构成战略优势。
Meta AI 近日发布 Brain2Qwerty v2,一种基于非侵入式脑磁图(MEG)信号的实时脑到文本解码管线,无需植入或手术。系统在 9 名志愿者约 22,000 句打字数据上训练(每人 10 小时),平均词准确率 61%(WER 39%),最高参与者达 78%,超过半数句子错误不超过一个词。管线由卷积编码器、Transformer 和字符级语言模型组成,并微调大语言模型提取语义表征,用 AI 智能体自动化改进解码流程但最终配置由人工选定。Meta 同时以 CC BY-NC 4.0 协议开源 v1 和 v2 训练代码。
Opinion from a former Meta PM. And this is from Aravind Srinivas of Perplexity "China can build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem." https://x.com/rohanpaul_ai/status/2071780740220752220/video/1
The worst case scenario for USA AI: 1. Chinese open sources keep gaining market share. China owns the model layer. 2. Those models were trained and inference-optimized on Huawei chips instead of NVIDIA. China also owns the chip layer. 3. US doesn't build data centers fast enough to keep up with the demand of compute, storage and energy. China meanwhile exports the inference and training layer(for continual training it will happen along with inference) Export control is not the right strategy here. Simply banning "open source from China" doesn't solve the issue here. USA must invest in open source models, hopefully get Chinese models to use NVIDIA, and invest in nuclear asap.
We’ve been impressed with GLM-5.2 and so are introducing a $9.99/month subscription to give you 2-5x discounted access to it and other open weight models like DeepSeek, Kimi, MiniMax, Mimo, Qwen. Use it on Cline CLI & IDE with $1.99 special promo if sign up via: npm i -g cline
Aravind Srinivas just explained why China’s open-source AI may become more powerful than ever. And why Anthropic has lobbied very hard for export control. "the only reason why there is even a 12-month gap between open source and frontier models is export controls. But there is a chance that, because of that, they now get really good at the physical layer. One advantage they (China) have is that they can actually build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem. And so, by forcing them to go out there and build all this, you are converting them into a far more potent competitor." --- From "20VC with Harry Stebbings" YouTube channel (@HarryStebbings ), link in comment
华为 openPangu-2.0-Flash 模型(总参数量 92B,激活参数量 6B)于 6 月 30 日正式开源上线,支持 512K 上下文。该模型为 openPangu 2.0 系列的两个版本之一,另一版本 Pro(505B 总参数,18B 激活参数)计划 7 月开源。openPangu 2.0 将从 6 月 30 日起陆续开源 7 大组件,包括模型权重、基础推理代码、训推算子等。
Imagine a lightweight AI that can read images AND chat with you. That's MiniCPM-V-4.6. It's a multimodal model that understands both text and visuals, perfect for on-device use. No heavy cloud needed.
美团今日发布万亿参数大模型 LongCat-2.0,总参数 1.6T,平均激活约 48B(动态 33B~56B),从零预训练,原生支持 1M 超长上下文,是业界首个在五万卡国产算力集群上完成全流程训练与推理的模型。预览版已在 OpenRouter 及 longcat.ai 开放,月调用量跻身 OpenRouter 全球前三。预训练数据超 30T tokens,通过 HCCL 异常处理等将月均日故障率降低 70% 以上,训练 MFU 提升 1.5 倍,稳态日吞吐超 1T tokens/day。推理采用 LSA 稀疏注意力和零计算专家实现 token 级动态激活。在 SWE-bench Pro 上获 59.5,领先 Gemini 3.1 Pro、GPT-5.5 和 Claude Opus 4.6;在其他多项评测中达到或接近前沿闭源模型水平。
👀 This is from an ex-Meta PM.
分析机构 SemiAnalysis 爆料,LeptonAI 创始人兼 CEO 贾扬清已从英伟达离职。2025 年 4 月,英伟达以 7 亿美元收购 LeptonAI,将贾扬清招入麾下。SemiAnalysis 称,英伟达曾承诺 2026 年开源 Lepton 核心软件平台,但 CEO 黄仁勋最终未批准。LeptonAI 创始团队来自阿里巴巴,曾用 500 行代码实现对话式搜索引擎演示,还推出端侧模型 Chrome 插件 Elmo 和提示词优化工具 PromptLLM。
AI News Radar 迎来大更新,新增自媒体板块,支持订阅某书、某音、某站、某X等平台账号,每日按热度推荐 Top10 信息(无热门则不硬凑),同时保留时间轴视图,可在热度优先和时间优先间切换。官方来源包括 OpenAI、Anthropic、Google 等一手消息及 GitHub AI&ML 更新日志。信息流按来源、类型、信号等级分类,标注高优先级与多源认证。项目完全开源,可零 API 部署独立 AI 日报页面,支持手机移动端及暗色界面。
推荐理由:我觉得这是目前最实用的AI信息过滤工具,新版直接整合了社媒热榜和官方源,内容创作者可以拿来当选题雷达,普通人也能一键部署自己的AI日报,省去每天刷各个平台的时间。
OpenClaw 推出免费 iOS 和 Android 配套节点应用,手机成为自托管智能体网络中的节点(非独立聊天机器人)。核心 Gateway 用 TypeScript 编写,需 Node 24 或 22.19+,运行在 macOS、Linux 或 Windows(WSL2)上。通过 WebSocket 端口 18789 与手机配对,支持本地 mDNS 发现或远程 Tailscale 连接。iOS 应用通过 QR 码配对,支持聊天、实时/后台语音模式、Canvas 表面,以及相机、屏幕、位置等设备能力。Android 应用提供流式聊天回复、图片附件、全会话历史、语音模式(ElevenLabs 或系统 TTS)和实时 Canvas。隐私敏感命令默认关闭,需在配置文件中显式允许。
Google Research 发布 TabFM,一种基于上下文学习(ICL)的零样本基础模型,专为表格数据的分类与回归任务设计。TabFM 将整个数据集(包含训练样本与待预测行)作为统一提示输入,在推理时直接从上下文中学习行列关系,无需模型训练、超参数调优或特征工程。与需大量人工调优的传统方法(如 XGBoost)不同,TabFM 通过单次前向传播即可在新表格上生成高质量预测。该模型已在 Hugging Face 和 GitHub 上开源。
推荐理由:TabFM 把表格预测变成一次前向传播,不再需调参和特征工程,数据科学家可以少熬夜了。BigQuery 集成让使用门槛降到 SQL,这是表格 ML 的零样本时刻。
Anthropic CEO Dario Amodei on Open-Source AI Models. "I don't think open source works the same way in AI that it has worked in other areas. Primarily because with open source you can see the source code of the model. Here we can't see inside the model, it's often called open weights instead of open source to kind of distinguish that. But a lot of the benefits, which is that many people can work on it and that it's kind of additive, don't quite work in the same way. So I've actually always seen it as a red herring. When I see a new model come out I don't care whether it's open source or not. If we talk about Deep Seek I don't think it mattered that Deep Seek is open source. I think I ask, is it a good model? Is it better than us at the things that matter? That's the only thing that I care about. It actually doesn't matter either way. Because ultimately you have to host it on the cloud. The people who host it on the cloud do inference. These are big models, they're hard to do inference on. When I think about competition I think about which models are good at the tasks that we do. I think open source is actually a red herring. It's not free. You have to run it on inference and someone has to make it fast on inference." --- From 'Alex Kantrowitz' YT channel (full video link in comment)
Anthropic CEO Dario Amodei on Open-Source AI Models. "I don't think open source works the same way in AI that it has worked in other areas. Primarily because with open source you can see the source code of the model. Here we can't see inside the model, it's often called open weights instead of open source to kind of distinguish that. But a lot of the benefits, which is that many people can work on it and that it's kind of additive, don't quite work in the same way. So I've actually always seen it as a red herring. When I see a new model come out I don't care whether it's open source or not. If we talk about Deep Seek I don't think it mattered that Deep Seek is open source. I think I ask, is it a good model? Is it better than us at the things that matter? That's the only thing that I care about. It actually doesn't matter either way. Because ultimately you have to host it on the cloud. The people who host it on the cloud do inference. These are big models, they're hard to do inference on. When I think about competition I think about which models are good at the tasks that we do. I think open source is actually a red herring. It's not free. You have to run it on inference and someone has to make it fast on inference." --- From 'Alex Kantrowitz' YT channel (full video link in comment)