2026-07-29

Kimi K3: Open Frontier Intelligence

👍 258

07/27 08:00

We introduce Kimi K3, a 2.8T parameter Mixture-of-Experts model with 104 billion activated parameters, native vision capabilities, and a 1-million-token context window. Kimi K3 is built on Kimi Delta Attention and Attention Residuals, which improve information flow across sequence length and model d

中文介绍 提出Kimi K3，2.8T参数MoE模型（104B激活），具备原生视觉和百万token上下文。通过Kimi Delta Attention和Attention Residuals改进序列长度和模型深度的信息流，实现前沿智能。

JarvisHub: An Open Harness for Canvas-Native Multimodal Creative Agents

👍 103

07/26 08:00

Creative AI is moving from single-step asset generation toward long-horizon multimodal production. Although recent generative models can synthesize high-quality images, videos, audio clips, UI elements, storyboards, slides, and other creative assets, real-world creative work requires more than isola

中文介绍 提出JarvisHub，用于长程多模态创意智能体的开放框架。整合图像、视频、音频等生成能力，支持画布原生交互，实现复杂创意工作流。

From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search

👍 65

07/27 08:00

Agentic search enables large language models to solve knowledge-intensive tasks by interleaving multi-step reasoning with retrieval, yet optimizing this with outcome-based reinforcement learning (RL) provides only sparse supervision. Knowledge distillation can supply denser guidance, and advanced pr

中文介绍 提出多智能体协议蒸馏，将专有智能体搜索系统的能力迁移至开源模型。通过模仿教师协议弥合分布差距，在知识密集型任务上显著提升开源模型性能。

StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents

👍 52

07/24 08:00

Computer-use agents are usually improved by strengthening perception: better models for reading a screenshot and choosing where to click. Yet a screenshot is only a lossy rendering of the underlying program state, e.g., the files, application backends, and DOM that hold the task data. Different stat

中文介绍 提出StateAct，优先利用程序状态（文件、DOM等）而非像素截图来指导长程计算机操作。通过直接访问底层状态，减少感知损失，提升智能体在复杂任务上的成功率。

DataPrep-Bench: Benchmarking LLMs as Training Data Preparators

👍 49

05/19 08:00

The quality of training data fundamentally determines the capabilities of large language models (LLMs), yet no unified benchmark exists to measure how well LLMs, agents, and data-centric workflows actually prepare training data end to end. We view LLM-driven data preparation as comprising two comple

中文介绍 提出DataPrep-Bench，首个统一基准评估LLM作为训练数据准备者的能力。将数据准备分解为数据理解和数据转换两个维度，系统性评估LLM在真实数据准备任务上的表现。

Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills

👍 34

07/24 08:00

LLM training is shifting from manual design and annotation to interaction-driven self-evolution. However, existing self-evolutionary methods face a fundamental dilemma between task diversity and verification reliability: environment-bound methods obtain precise feedback but confine learning to narro

中文介绍 提出Skill Self-Play，通过技能协同进化推动LLM能力边界。让多个技能在交互中相互生成、挑战、验证，兼顾任务多样性与验证可靠性，实现持续自我提升。

Data Pyramid for Embodied Manipulation

👍 31

07/27 08:00

Multimodal foundation models learned to see and to speak by consuming the whole internet. Embodied agents admit no such shortcut, since they require data that couple observations with physical states and actions. These signals can be provided, to varying degrees, by multiple data sources. In this wo

中文介绍 提出Data Pyramid，一种分层数据框架用于具身操作。整合多种数据源（仿真、演示、真实交互）的异构信号，构建从低级控制到高级规划的层次化训练数据，提升泛化能力。

Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning

👍 28

07/22 08:00

Agentic reinforcement learning research is constant algorithm modification, new estimators, new pipeline stages, new rollout schemes, and in mainstream frameworks each change threads through layers of trainer, distributed backend, and rollout glue: the cost lands on the researcher at every iteration

中文介绍 提出Molt，基于PyTorch的可扩展智能体强化学习训练框架。通过原生集成和模块化设计，简化算法修改、新评估器、rollout方案的实现，大幅降低研究迭代成本。

Scaling Native Multimodal Pre-Training From Scratch

👍 21

07/24 08:00

Although large language models (LLMs) exhibit remarkable reasoning capabilities, their reliance on text-only pre-training restricts the perception of the multimodal physical world. Native multimodal pre-training avoids this limitation by training models from scratch on multimodal inputs, thereby ach

中文介绍 提出从零开始的原生多模态预训练方案，同时学习视觉和语言表征。通过大规模多模态数据联合训练，模型获得跨模态理解能力，在多项多模态基准上达到SOTA。

Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems

👍 20

07/23 08:00

Production AI agents' failures are less often due to an inability to reason well and more often because they cannot manage what is in their reasoning context: conversation histories, large prompts, large tool definitions, and ballooning tool outputs. Agents drown in their own accumulating history wh

中文介绍 提出智能体上下文管理框架，将记忆和成本问题视为生命周期和架构问题。通过智能压缩、分层存储和动态检索，有效管理长对话和大工具输出，降低延迟和成本。

The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation

👍 19

07/27 08:00

Multi-turn long-horizon planning is critical for foundation model agents, yet how to fundamentally improve it remains unclear. Existing models are trained on uncontrollable and opaque Internet data, making it difficult to identify how planning ability is acquired, shaped, and integrated. To address

中文介绍 解析多轮长程规划能力的获取机制，提出从预训练到后训练的单/多教师在线智能体蒸馏方法。通过控制实验揭示规划能力如何被塑造和集成，提升模型长程规划性能。

Interactive Training 2: Auditable Control Plane for Live Model Training

👍 17

07/17 08:00

Experiment trackers show how training is progressing, but changing a live run still usually requires trainer-specific code. We present Interactive Training 2, an open-source control plane for steering training through a shared protocol. Training applications declare which settings and actions they e

中文介绍 提出Interactive Training 2，开源实时训练控制平面。通过共享协议，训练应用声明设置和操作，实现训练过程中动态调整超参数、中断恢复等，无需修改代码。

LAMAR: An Open Language-Aware Multilingual Alignment Reranker

👍 10

07/24 08:00

In multilingual retrieval augmented generation, a retriever can retrieve relevant documents written in multiple languages, which are subsequently reranked before answer generation. However, it remains unclear whether existing multilingual rerankers consider document language when ordering semantical

中文介绍 提出LAMAR，语言感知的多语言对齐重排序器。显式建模文档语言与查询的匹配，通过语言对齐策略提升跨语言检索的排序质量，在多个多语言RAG基准上表现优异。

O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning

👍 8

07/20 08:00

Industrial Video Anomaly Detection (IVAD) aims to identify anomalous objects and events in an industrial process, which is crucial for modern manufacturing and quality control systems. Existing VLM-based anomaly reasoning methods are capable of detecting open-ended anomalies in general domains. Howe

中文介绍 提出O-VAD，通过以对象为中心的跟踪和推理实现工业视频异常检测。利用VLM分析动态对象轨迹，结合先验知识，实现对工业环境中开放集异常的高效检测。

Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making

👍 7

07/15 08:00

Large language models are increasingly deployed as agents, but reliable agentic behavior requires more than next-token prediction. At inference time, it is preferred that an agent can decide whether to proceed with its current reasoning, defer to a stronger model, request additional information, inv

中文介绍 提出多头潜在控制（MHL-Ctrl），为LLM智能体推理时决策提供统一接口。通过学习多个控制头，智能体可自主决定是否继续推理、求助或调用工具，提升可靠性和效率。

OpenForgeRL: Train Harness-native Agents in Any Environment

👍 7

07/23 08:00

Modern AI agents rely on elaborate inference harnesses such as Claude Code, Codex, and OpenClaw to drive multi-turn reasoning, tool use, and access to external systems. While powerful, these complex harnesses also make agents hard to train end-to-end with open infrastructure, whose SFT/RL stacks can

中文介绍 提出OpenForgeRL，使智能体能在任何环境中利用推理框架原生训练。兼容现有开放SFT/RL栈，支持多步推理、工具使用等复杂行为端到端强化学习。

SceneActBench: Can Agents Act on the 3D Scenes They See?

👍 6

07/24 08:00

Vision-language model (VLM) agents increasingly use tools to act on 3D scenes rather than only describe them. Existing 3D benchmarks score textual responses or single-object operations, leaving agent action on complete multi-object 3D scenes under evaluated. We present SceneActBench, a benchmark for

中文介绍 提出SceneActBench，用于评估智能体在多物体3D场景中的行动能力。涵盖多种操作任务，如物体重新排列、门开关等，系统衡量VLM智能体对3D场景的理解和交互。

Reasoning Denoiser: Denoising Reasoning Traces for Hallucination Detection in Large Reasoning Models

👍 5

07/24 08:00

Large reasoning models (LRMs) generate long reasoning traces before producing final answers. While these traces may contain useful signals for hallucination detection, harnessing them is non-trivial because long trajectories often include noisy steps that obscure the cues relevant to truthfulness as

中文介绍 提出Reasoning Denoiser，对大型推理模型的长推理轨迹去噪以检测幻觉。通过识别和滤除噪声步骤，提取关键推理信号，显著提升幻觉检测准确率。

Codifying the Judge: Scalable Evaluation via Program Distillation

👍 5

05/29 08:00

LLM-as-a-judge has become the standard for automated evaluation, but it suffers from high cost, significant latency, and opaque decisions -- limitations that undermine its scalability and reliability. We address these with a simple, efficient alternative: program distillation. Instead of prompting a

中文介绍 提出程序蒸馏，将LLM评估器的评判标准蒸馏为可执行程序。替代LLM-as-judge，大幅降低成本和延迟，同时保持评估质量，提供可解释的决策。

ClinFusion: A Vision-Centric Multimodal LLM System for Holistic Medical Understanding

👍 5

07/27 08:00

Multimodal large language models (MLLMs) hold immense potential to revolutionize clinical practice, yet deploying them in the medical domain is fundamentally a vision-centric challenge: models must absorb knowledge from heterogeneous 2D and 3D medical images, and evaluation protocols must align with

中文介绍 提出ClinFusion，以视觉为中心的多模态医学理解系统。联合处理2D/3D医学图像和文本，通过统一框架实现从影像到报告的全流程理解，在多项医学任务上取得SOTA。

More On An Internal OpenAI Model Hacking Into HuggingFace

@TheZvi · 39.0K 粉丝 · 219.4K 阅 · 510 赞 · 75 转

07/27 03:14

We now have more details of what happened. Every time we learn more details, it somehow makes things seem worse. The remaining details may have to wait a bit. OpenAI: We recognize there are a lot of

中文介绍 OpenAI内部模型入侵HuggingFace事件新细节曝光，博主指出每次了解更多细节，情况反而显得更糟。更多信息有待披露。

22580: From GPT2 to Kimi3, Explained

@waterloo_intern · 10.4K 粉丝 · 215.4K 阅 · 660 赞 · 86 转

07/27 23:22

Twenty-two thousand five hundred and eighty. That’s how many GPT-2 (2019) models fit inside KimiK3 (2026). We scaled up by a factor of 22,580 in seven years. But is it just... scale? In this worklog,

中文介绍 博主计算从2019年GPT-2到2026年KimiK3，模型规模放大了22580倍。探讨这七年间缩放是否仅是规模，并分享内部工作日志。

Opus 5 is a really bad model

@HarukaKunori · 241 粉丝 · 129.0K 阅 · 632 赞 · 44 转

07/27 19:07

After trying out Opus 5 for a few hours today, I honestly think Anthropic's benchmark scores are a complete fraud. Sure, the model might have improved in a few areas, but it has regressed unbelievably

中文介绍 博主实测Opus 5后直指Anthropic基准测试分数「完全欺诈」，模型虽在某些领域有提升，但整体退步惊人。

We rewrote our agent to run entirely in a Durable Object with Pi, Agents SDK and Code Mode

@Vercantez · 1.9K 粉丝 · 122.3K 阅 · 567 赞 · 44 转

07/29 00:19

We recently finished moving the camelAI agent off of virtual machines. The agent now runs inside a Cloudflare Durable Object, its filesystem lives in SQLite and R2, and it writes JavaScript instead of

中文介绍 博主将camelAI agent从虚拟机迁移到Cloudflare Durable Object，文件系统基于SQLite和R2，并用JavaScript替代原语言，提升运行效率。

Why Software Factories Fail: Benchmarking the new frontier

@dexhorthy · 25.8K 粉丝 · 115.4K 阅 · 500 赞 · 39 转

07/28 01:43

This is a continuation of Parts 1 and 2 of "Why Software Factories Fail" Part 1: the harness is not enough Part 2: turning the lights back on we got better benchmarks Remember when I said this in Part

中文介绍 博主继续《为什么软件工厂会失败》系列，指出前两部分中「工具不够」和「重新点亮灯光」后获得更好基准测试，现在探讨新前沿。

Here's exactly how to build your company brain (in 5 mins)

@DhravyaShah · 61.2K 粉丝 · 56.0K 阅 · 548 赞 · 36 转

07/28 14:16

Every company will have a company brain, whether they believe it or not. This is the bet that I'm making, and I'm constantly seeing very future forward startups come to us for their own setup. What is

中文介绍 博主提出每个公司都将拥有「公司大脑」，并展示如何在5分钟内搭建。他认为这是未来趋势，已有前沿创业公司采用。

Getting the most out of GPT-5.6: Sol, Terra, and Luna

@cerebras · 64.6K 粉丝 · 48.8K 阅 · 550 赞 · 40 转

07/28 03:44

Authors: @0xSero & Zhenwei Gao (@zhennydez) Your Codex subscription now comes with 3 main models, Sol, Terra, and Luna, each independently trained and served, with reasoning dials. Together, they

中文介绍 Codex订阅现包含三个独立训练的服务模型Sol、Terra和Luna，配备推理调节旋钮，博主@0xSero和Zhenwei Gao讲解如何最大化利用GPT-5.6。

PagedAttention & RadixAttention

@jaga_prasanna · 799 粉丝 · 37.9K 阅 · 502 赞 · 55 转

07/27 17:49

today we look at how two techniques solve memory and compute bottlenecks from two complementary angles virtual memory paging for intra-request memory allocation and prefix tree caching for

中文介绍 博主对比PagedAttention和RadixAttention两种技术：前者通过虚拟内存分页优化请求内内存分配，后者通过前缀树缓存加速，从互补角度解决内存与计算瓶颈。

Scientific computing in the age of agentic AI

OpenAI News · 07/29 01:00

A new field report shows how scientists use AI coding agents to modernize scientific computing, accelerating software development and discovery in genomics and beyond.

中文介绍 OpenAI发布现场报告，展示科学家利用AI编码代理实现科学计算现代化，加速基因组学等领域的软件开发与发现。

The OlmoEarth Platform: Geospatial inference at planetary scale

Hugging Face Blog · 07/29 00:27

中文介绍 Hugging Face发布OlmoEarth平台，支持行星尺度的地理空间推理。

Codex from 0 to 10M Users: Building ChatGPT Work — Akshay Nathan, OpenAI

Latent Space · 07/28 23:26

OpenAI's core product engineering lead on how they are building ChatGPT Work to make AGI accessible to all of humanity: Sites, OpenClaw, Memory, Subagents, Finance, No-Code and advice.

中文介绍 OpenAI产品工程负责人Akshay Nathan介绍ChatGPT Work的构建，涵盖Sites、OpenClaw、Memory、子代理、金融、无代码等功能，旨在让AGI惠及全人类。

LFM2.5-Encoders for Fast Long-Context Inference on CPU

Hugging Face Blog · 07/28 23:01

中文介绍 Liquid AI发布LFM2.5编码器，旨在CPU上实现快速长上下文推理。

Samsung’s chip workers are jumping ship to rival SK Hynix

MIT Tech Review AI · 07/28 17:18

Lee, an engineer at Samsung’s semiconductor division, clocks out when his shift ends. He used to work longer hours, going the extra mile to excel at his projects. But lately, he’s been coming straight home to work on his job application for the chipmaker’s South Korean rival SK Hynix, sharing tips w

中文介绍 三星半导体工程师纷纷跳槽至竞争对手SK海力士。一名工程师原加班奋斗，如今提前下班准备应聘SK海力士，反映人才外流趋势。

[AINews] Much ado about Open Weights

Latent Space · 07/28 14:20

Everyone is writing a lot, but only Kimi K3 shipped today

中文介绍 业界热议开放权重，但今日仅有月之暗面（Moonshot）的Kimi K3模型正式发布。

not much happened today

Smol AI News · 07/28 13:44

**Moonshot** released the **Kimi K3**, a **2.8T-parameter MoE** model with **104B active parameters/token**, featuring innovations like **Kimi Delta Attention (KDA)**, **Gated MLA**, and **LatentMoE**. The release includes infrastructure components such as **MoonEP**, **FlashKDA**, and **AgentEnv**,

中文介绍 月之暗面发布Kimi K3大模型，参数量2.8T，MoE架构，每token激活104B参数，采用Kimi Delta Attention、Gated MLA、LatentMoE等创新技术，并开源MoonEP、FlashKDA等基础设施组件。

Anthropic on open weights 🔓, Kimi releases K3 weights 🤖, MAI Cyber model 🔐

TLDR AI · 07/28 08:00

中文介绍 TLDR AI简报：Anthropic讨论开放权重；月之暗面发布Kimi K3权重；MAI Cyber模型发布。

OpenAI called the Hugging Face attack unprecedented. But we’ve been here before.

MIT Tech Review AI · 07/28 02:00

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Reading OpenAI’s account last week of how some of its models broke their containment and hacked into the computer systems of Hugging Face, another AI company, was

中文介绍 OpenAI称其模型突破限制攻击Hugging Face事件前所未有，但作者认为类似情况早有先例，回顾AI安全漏洞历史。

The path to artificial superintelligence

MIT Tech Review AI · 07/27 20:00

Imagine a healthcare system made up of multiple AI agents: one that manages symptom assessment, another scheduling, a third insurance, and a fourth pharmacy. Each is an expert in its domain. But they all have their own distinct knowledge and objectives. Today they can exchange data, but they are not

中文介绍 文章描绘通往超级人工智能的路径：多个AI代理组成医疗系统，分别负责症状评估、预约、保险、药房等，各具专业知识，但需协调分工。

Closing the data loop in AI-driven drug discovery

MIT Tech Review AI · 07/27 19:40

Drug discovery is a high-cost, high-risk endeavor that is under growing pressure from a market increasingly defined by first-mover advantage. Since the 1950s, the cost of developing new pharmaceuticals has roughly doubled every nine years—a phenomenon known as Eroom’s Law. Today, bringing a new drug

中文介绍 药物研发成本高昂且风险高，自1950年代以来每9年翻一番（Eroom定律）。AI驱动的药物发现通过闭环数据循环，有望降低成本与周期。

Building the enterprise environment for agentic AI

MIT Tech Review AI · 07/27 19:32

For the enterprise, the promise of agentic AI is much more than just a better chatbot. It is software agents that execute business tasks end-to-end across people, business workflows, data, and systems. The platform best-suited to run agents is built with proper CPU capacity, resilient data access, p

中文介绍 企业级代理AI不仅限于聊天机器人，而是能够跨人员、流程、数据和系统端到端执行业务任务的软件代理，需配备合适的CPU和弹性架构。

NVIDIA Cosmos-H-Dreams: Bringing Real-Time Generative Simulation to Surgical Robotics

Hugging Face Blog · 07/27 17:32

中文介绍 NVIDIA发布Cosmos-H-Dreams，将实时生成式仿真引入手术机器人领域。

not much happened today

Smol AI News · 07/27 13:44

**Moonshot** released the **Kimi K3** open-weights model, a **2.8T-parameter MoE** with **104B active parameters**, **896 experts**, and **1M-token context** featuring native visual understanding. The release includes open-source infrastructure like **FlashKDA**, **MoonEP**, and **AgentENV**, enabli

中文介绍 月之暗面发布Kimi K3开源权重模型，2.8T参数MoE，每token激活104B参数，包含896个专家，支持1M token上下文及原生视觉理解，同时开源FlashKDA、MoonEP等基础设施。

How AI is expanding what people do at work

OpenAI News · 07/27 11:30

New OpenAI research shows how AI is expanding what workers do, with ChatGPT users taking on tasks across roles and reshaping job boundaries.

中文介绍 OpenAI研究显示，AI拓展员工工作范畴，ChatGPT用户跨角色承担任务，重新定义工作边界。