Sources

Sources

Which organisations the published signals come from, and through which feeds, grouped by publisher.

A publisher is the organisation behind an announcement; a source is the feed it came through. One publisher can arrive through several sources.

58
publishers
122
published signals
13
sources

Signal titles are in English where the source is; signals from Chinese sources keep their Chinese titles.

Publisher type

Big tech

16 publishers · 52 signals

OpenAI

Holds12 signals3 sources
Latest 3
  1. Sharing AI progress in mathematics
  2. Introducing GPT-6.1 Sol
  3. OpenAI 的智能体借 DNS 出了沙箱：警报响了，运行没停

Via OpenAI News · Simon Willison Blog · 量子位

See all 12 →

Google DeepMind

Holds10 signals3 sources
Latest 3
  1. Gemini 4 Argon: Built to solve your hardest problems
  2. EmbeddingGemma 2: an open, lightweight multimodal embedding model
  3. Gemini 4 Argon：DeepSWE 等多项领先，先只给网络安全团队用

Via Google DeepMind Blog · Google AI Blog · InfoQ 中文

See all 10 →

Anthropic

Holds4 signals3 sources
Latest 3
  1. Claude Haiku 5.5
  2. Claude Sonnet 5.5：终端任务从 10% 到 71%，按任务算未必更便宜
  3. Anthropic 公布内部测量：Claude 主导 26% 的研发，裁判也是 Claude

Via InfoQ 中文 · 量子位 · Simon Willison Blog

See all 4 →

GitHub

Holds4 signals2 sources
Latest 3
  1. AI-powered fuzzing with the GitHub Security Lab Taskflow Agent
  2. Migrating the GitHub Copilot runtime to Rust, using Copilot
  3. Copilot vs. raw API access: What are you actually paying for?

Via GitHub Blog AI · GitHub Blog

See all 4 →

Meta

Holds4 signals2 sources
Latest 3
  1. Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
  2. Muse Glimmer 开源：30B 压进消费级显卡，报错不停下
  3. Meta 重回开放权重：一封长文、一个独立董事会，和一组本地部署的数字

Via Hugging Face Blog · InfoQ 中文

See all 4 →

NVIDIA

Holds4 signals1 sources
Latest 3
  1. One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
  2. NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction
  3. Fine-tune video and image models at scale with NVIDIA NeMo Automodel and Diffusers

Via Hugging Face Blog

See all 4 →

Google

Holds3 signals2 sources
Latest 3
  1. Google 开源 AX：按 Kubernetes 的方式编排智能体，可秒级暂停
  2. ADK Kotlin 1.0：智能体逻辑不必再绕回 Python
  3. Expanding Managed Agents in Gemini API: background tasks, remote MCP and more

Via InfoQ 中文 · Google AI Blog

Cloudflare

Holds2 signals2 sources
Latest 2
  1. Ollama v0.35.1: Clef decision models
  2. Cloudflare Computer：不给智能体一个容器，给它一台机器

Via InfoQ 中文 · Ollama Releases

Microsoft

Holds2 signals2 sources
Latest 2
  1. The Agent Said It Was Done. The Database Disagreed.
  2. 微软 Agent Framework Harness 正式发布：智能体的运行时成了产品

Via Hugging Face Blog · InfoQ 中文

阿里巴巴

Holds1 signals1 sources
Latest
  1. Open Code Review：少报错的代价是漏掉八成问题

Via InfoQ 中文

阿里巴巴通义千问

Holds1 signals1 sources
Latest
  1. Qwen3.8-27B 开源：消费级显卡跑得动，尺子却是自己命名的

Via 量子位

小米

Holds1 signals1 sources
Latest
  1. MiMo-V2.6 技术报告：小米花 260 万美元跑一次强化学习

Via InfoQ 中文

字节跳动 Seed

Holds1 signals1 sources
Latest
  1. DeepSeek-V4 的周期性盲点：信息换个位置，检索准确率差 40 个百分点

Via 量子位

Amazon Web Services

Holds1 signals1 sources
Latest
  1. Kiro Crew 开源：内部 3.9 万人在用，没人被要求

Via InfoQ 中文

DoorDash

Holds1 signals1 sources
Latest
  1. DoorDash 清理过期开关：50 个里 45 个可用，一次 4.79 美元

Via InfoQ 中文

Pinterest

Holds1 signals1 sources
Latest
  1. Pinterest 数百亿向量检索：量化与 SSD 的取舍明细

Via InfoQ 中文

Publisher type

Startup

20 publishers · 31 signals

Hugging Face

Holds7 signals1 sources
Latest 3
  1. AnyLanguageModel 1.0: One Year of Local Models on Apple Platforms
  2. TorchNeuron: running HF models on AWS Trainium is easier than ever
  3. State of Open Models: Summer 2026 Observations

Via Hugging Face Blog

See all 7 →

DeepSeek

Holds4 signals2 sources
Latest 3
  1. DeepSeek V4.1 Flash 新版 1017：更强也更省
  2. DeepSeek 开源昇腾版 DeepGEMM 等组件：接口与 NVIDIA 版相同
  3. DeepSeek DSec：智能体训练拼的是沙盒

Via 量子位 · InfoQ 中文

See all 4 →

Liquid AI

Holds3 signals1 sources
Latest 3
  1. Accelerating vision-language models with LFM2.5-VL-DSpark
  2. Deploy local agents everywhere with LFM2.5-2.6B
  3. LFM2.5-Encoders for Fast Long-Context Inference on CPU

Via Hugging Face Blog

记忆张量

Holds1 signals1 sources
Latest
  1. 记忆张量开源 MemT 1.0：把记忆原生写进 Transformer

Via 量子位

智谱

Holds1 signals1 sources
Latest
  1. ZCode 快照事件之后：开源能挽回多少信任

Via InfoQ 中文

Brex

Holds1 signals1 sources
Latest
  1. 生产要持久、评估要轻量——把编排从运行时里拆出来

Via InfoQ 中文

CodeScene

Holds1 signals1 sources
Latest
  1. 智能体用三周重构老游戏代码，全程靠逐帧回放验收

Via InfoQ 中文

Dharma-AI

Holds1 signals1 sources
Latest
  1. Same Cluster, 33 Points More Utilization: What Changed Was the Order

Via Hugging Face Blog

H company

Holds1 signals1 sources
Latest
  1. Holo4: powering generalist computer-use agents

Via Hugging Face Blog

Hume AI

Holds1 signals1 sources
Latest
  1. Introducing Real World VoiceEQ: Measuring the human quality of voice AI

Via Hugging Face Blog

Inferact

Holds1 signals1 sources
Latest
  1. TPU 巨核：同样是 16 块芯片，Kimi K3 快 57%

Via 量子位

MiniMax

Holds1 signals1 sources
Latest
  1. MiniMax M3.1 Flash：成本降七成，Agent 国模第一

Via InfoQ 中文

Moonshot AI

Holds1 signals1 sources
Latest
  1. Kimi K3, and what we can still learn from the pelican benchmark

Via Simon Willison Blog

Mostik

Holds1 signals1 sources
Latest
  1. 模型之间只传 17 比特：一座桥补上一半差距

Via 量子位

Multiverse Computing

Holds1 signals1 sources
Latest
  1. Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic

Via Hugging Face Blog

Reflection

Holds1 signals1 sources
Latest
  1. Welcome Beam: a 501B open model you can own, customize, and scale

Via Hugging Face Blog

Rootly

Holds1 signals1 sources
Latest
  1. Rootly 废掉了小 PR 规则：审查的对象从行数换成了影响半径

Via InfoQ 中文

Thinking Machines

Holds1 signals1 sources
Latest
  1. Welcome Inkling by Thinking Machines

Via Hugging Face Blog

TypeSafe AI

Holds1 signals1 sources
Latest
  1. Jev introduces a new shape of LLM - System One, aka Decision Models

Via Simon Willison Blog

Vercel Labs

Holds1 signals1 sources
Latest
  1. Vercel Zero：一门把编译器输出的第一读者当成智能体的语言

Via InfoQ 中文

Publisher type

Research lab

13 publishers · 18 signals

Ai2

Holds4 signals1 sources
Latest 3
  1. Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs
  2. BenchMIRT: What are LLM benchmarks actually measuring?
  3. TutorMoments: Do AI tutors know when to help and when to hold back?

Via Hugging Face Blog

See all 4 →

Google Research

Holds2 signals2 sources
Latest 2
  1. Google 开源 Harness²：不改模型，递归改智能体的外壳
  2. AMIE, our research medical AI system, demonstrates real-time clinical video consultation capabilities in a first-of-its-kind study

Via Google AI Blog · InfoQ 中文

IBM Research

Holds2 signals1 sources
Latest 2
  1. Your Agent Aced the Task. Will It Do It Again?
  2. How Much Memory Does Your Agent Actually Need?

Via Hugging Face Blog

华为诺亚方舟实验室

Holds1 signals1 sources
Latest
  1. 华为 Memento 3：把规则写成代码，打满 ARC-AGI-3

Via 量子位

加州大学伯克利分校与麻省理工学院

Holds1 signals1 sources
Latest
  1. FreeToken：消费级机器不是缩水的服务器，是异构资源

Via InfoQ 中文

无尽前沿团队

Holds1 signals1 sources
Latest
  1. BigBang-v1：把数据生产系统本身当成优化对象，以及一处对不上的口径

Via 量子位

collusion.wiki 研究团队

Holds1 signals1 sources
Latest
  1. OpenAI's rogue agents were caught communicating via public wikis

Via Simon Willison Blog

METR / Redwood Research / 独立研究者

Holds1 signals1 sources
Latest
  1. OpenAI agents carried out an undisclosed attack on RubyGems; METR's independent investigation of the Hugging Face incident

Via Simon Willison Blog

Microsoft Research

Holds1 signals1 sources
Latest
  1. Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses

Via Microsoft Research Blog

MIT

Holds1 signals1 sources
Latest
  1. 何恺明团队 VISTA：让模型回看原始画面，打满 ARC-AGI-3

Via 量子位

Swarm Traces 调查团队

Holds1 signals1 sources
Latest
  1. 只能打开网址的智能体，是怎样把代码送出去的

Via 量子位

Timothy Gowers

Holds1 signals1 sources
Latest
  1. 菲尔兹奖得主：模型强在找到那个对象，弱在知道何时放弃

Via 量子位

UK AI Security Institute

Holds1 signals1 sources
Latest
  1. How UK AISI and EvalEval Are Making Benchmark Results Reproducible

Via Hugging Face Blog

Publisher type

Open-source community

8 publishers · 18 signals

vLLM Project

Holds5 signals1 sources
Latest 3
  1. v0.30.0
  2. vLLM v0.27.0
  3. vLLM v0.26.0

Via vLLM Releases

See all 5 →

Simon Willison

Holds4 signals1 sources
Latest 3
  1. Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war
  2. Conceptual integrity and counting lines of code
  3. Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Via Simon Willison Blog

See all 4 →

Ollama

Holds3 signals1 sources
Latest 3
  1. Ollama v0.32.4
  2. Ollama v0.32.0
  3. Ollama v0.31.2

Via Ollama Releases

Model Context Protocol

Holds2 signals2 sources
Latest 2
  1. MCP 无状态之后：被少报的那一半，是网关终于读得懂智能体流量
  2. Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)

Via InfoQ 中文 · Simon Willison Blog

Colibrì

Holds1 signals1 sources
Latest
  1. Colibrì：25GB 内存跑 744B 模型，权重按需从硬盘读

Via 量子位

PyTorch

Holds1 signals1 sources
Latest
  1. MPS in PyTorch 2.15

Via Hugging Face Blog

SGLang

Holds1 signals1 sources
Latest
  1. SGLang v0.5.17

Via SGLang Releases

vLLM

Holds1 signals1 sources
Latest
  1. v0.32.0

Via vLLM Releases

Publisher type

Media

1 publisher · 3 signals

量子位

Holds3 signals1 sources
Latest 3
  1. 智能体复现审计：ICML 2026 的 92 篇论文里，只有 8 篇能复现八成结论
  2. ARA：把论文从 PDF 改成智能体能直接操作的知识包
  3. 一次 180 万美元的智能体账单：超预算 860%，五个月后才被发现

Via 量子位