New AI models, open-source releases, research and trending projects, one plain line each, linked to the original source. Updated daily at 7 PM ET.https://whatisai.fyihttps://github.com/cloudflare/security-audit-skillhttps://github.com/cloudflare/security-audit-skillCloudflare's skill for coding agents that runs a security review of a codebase in six phases and checks each finding before writing it to a report.Sat, 24 Oct 2026 12:00:00 GMTGitHub trendingcloudflarehttps://github.com/NVIDIA/OpenShellhttps://github.com/NVIDIA/OpenShellNVIDIA's runtime that lets fleets of AI agents read files, call APIs and use credentials inside locked-down sandboxes, without open access to your data or network.Sat, 24 Oct 2026 12:00:00 GMTGitHub trendingNVIDIAhttps://github.com/pbakaus/impeccablehttps://github.com/pbakaus/impeccableDesign guidance for AI coding agents, with commands and automatic checks that catch the usual look of AI-made web pages, such as purple-to-blue gradients.Sat, 24 Oct 2026 12:00:00 GMTGitHub trendingpbakaushttps://github.com/Nutlope/hallmarkhttps://github.com/Nutlope/hallmarkA skill from Together AI that helps coding agents such as Claude Code and Codex design web pages, then tests them in a real browser for habits that look machine-made.Sat, 24 Oct 2026 12:00:00 GMTGitHub trendingNutlopehttps://github.com/makecindy/cindyhttps://github.com/makecindy/cindyAn open-source AI agent app for desktop and phone that runs on your own computer and can switch between Claude Code, Codex and different models in the middle of a task.Sat, 24 Oct 2026 12:00:00 GMTGitHub trendingmakecindyhttps://github.com/luongnv89/claude-howtohttps://github.com/luongnv89/claude-howtoA visual guide to Claude Code, Anthropic's coding agent, that goes from the basics to advanced agents with examples and templates you can copy and paste.Sat, 24 Oct 2026 12:00:00 GMTGitHub trendingluongnv89https://github.com/pollen-robotics/microduckhttps://github.com/pollen-robotics/microduckThe software brain of Microduck, a walking duck robot of roughly 25 cm from Pollen Robotics that moves using neural networks trained by reinforcement learning (trial and error).Sat, 24 Oct 2026 12:00:00 GMTGitHub trendingpollen-roboticshttps://github.com/google-gemini/gemini-clihttps://github.com/google-gemini/gemini-cliGoogle's open-source AI agent that runs Gemini in your terminal. Google now points people to its newer Antigravity CLI for the latest Gemini models.Sat, 24 Oct 2026 12:00:00 GMTGitHub trendinggoogle-geminihttps://github.com/aws/agent-toolkit-for-awshttps://github.com/aws/agent-toolkit-for-awsOfficial AWS add-ons for coding agents such as Claude Code, Codex and Cursor, giving them the tools and guardrails to build, deploy and manage apps on AWS.Sat, 24 Oct 2026 12:00:00 GMTGitHub trendingawshttps://www.kimi.com/en/blog/kimi-k3-1-prohttps://www.kimi.com/en/blog/kimi-k3-1-proMoonshot says K3.1 Pro can read up to 1 million tokens (word pieces) at once and nearly matches Claude Opus 5.5 on a coding test. It is out now in the Kimi apps and API, with the weights promised by 4 November.Fri, 23 Oct 2026 12:00:00 GMTModelsMoonshot AIhttps://block.xyz/dessahttps://block.xyz/dessaBlock says it trained Dessa from scratch on human conversation alone, with no text and no pretrained models, so it can hold a live spoken chat full of pauses and emotion. Block has opened a waitlist.Fri, 23 Oct 2026 12:00:00 GMTModelsBlockhttps://github.com/google/adk-python/releases/tag/v2.13.0https://github.com/google/adk-python/releases/tag/v2.13.0The Python kit for building AI agents moves 12 experimental features to stable, adds real-time voice chats that users can interrupt, and can keep sessions and long-term memory in a MongoDB database.Fri, 23 Oct 2026 12:00:00 GMTOpen sourceGooglehttps://research.google/blog/bringing-foundational-models-to-the-front-lines-of-tb-screening/https://research.google/blog/bringing-foundational-models-to-the-front-lines-of-tb-screening/Google says the chest X-ray screening tool that grew from its research is used in about 40 health facilities in six countries, and that miners are checked on site four kilometres underground.Fri, 23 Oct 2026 12:00:00 GMTResearchGoogle Researchhttps://corgi.computer/https://corgi.computer/Like tmux, Corgi runs many terminal sessions in one window, and it shows whether each agent, such as Claude Code or Codex, is working, finished or waiting for you. It runs on macOS, Linux and WSL.Fri, 23 Oct 2026 12:00:00 GMTTrendingCorgihttps://github.com/anthropics/knowledge-work-pluginshttps://github.com/anthropics/knowledge-work-pluginsAnthropic's open-source plugins for Claude Cowork, built mainly to help knowledge workers, people whose job is mostly working with information, get everyday tasks done.Fri, 23 Oct 2026 12:00:00 GMTGitHub trendinganthropicshttps://github.com/SnailSploit/Claude-Redhttps://github.com/SnailSploit/Claude-RedA set of skill files that turn Claude into a specialist for authorised security testing, such as red-team exercises and bug bounty hunts, each loading only when a conversation needs it.Fri, 23 Oct 2026 12:00:00 GMTGitHub trendingSnailSploithttps://github.com/OpenBMB/VoxCPMhttps://github.com/OpenBMB/VoxCPMOpenBMB's text-to-speech system; version 2 has 2 billion parameters, speaks 30 languages, can design new voices or clone one, and outputs 48 kHz audio.Fri, 23 Oct 2026 12:00:00 GMTGitHub trendingOpenBMBhttps://developers.openai.com/api/docs/guides/decisionshttps://developers.openai.com/api/docs/guides/decisionsThe API reads text, images or JSON and returns a fixed-form answer, such as the chance something is true, one pick from a list or a score, which OpenAI says comes about 10 times faster than its Responses API.Thu, 22 Oct 2026 12:00:00 GMTModelsOpenAIhttps://huggingface.co/internlm/Atria-Dawnhttps://huggingface.co/internlm/Atria-DawnBuilt on GLM-5.2, a 744-billion-parameter model, Atria Dawn targets research and engineering jobs that take many steps and tools, and aims for results others can check. It uses the MIT licence.Thu, 22 Oct 2026 12:00:00 GMTOpen sourceAtriahttps://huggingface.co/tencent/WeVisOne-VL-4Bhttps://huggingface.co/tencent/WeVisOne-VL-4BBuilt on Qwen3-VL, the 2B and 4B models answer requests about a picture by drawing boxes, pointing, counting, reading text or tracing an object's exact outline. They use the Apache 2.0 licence.Thu, 22 Oct 2026 12:00:00 GMTOpen sourceTencenthttps://huggingface.co/blog/allenai/olmoearth-v1-3https://huggingface.co/blog/allenai/olmoearth-v1-3The model turns a year of satellite images into 128 numbers, called an embedding, for every 10-metre patch of Earth. Ai2 shares these for the whole planet, plus the weights, training data and code.Thu, 22 Oct 2026 12:00:00 GMTOpen sourceAi2https://www.microsoft.com/en-us/research/blog/flint-interactive-a-language-for-creating-interactive-data-visualizations/https://www.microsoft.com/en-us/research/blog/flint-interactive-a-language-for-creating-interactive-data-visualizations/Microsoft Research says the language lets people and agents describe what a click or drag should do, such as highlighting linked points, while its compiler takes care of the low-level chart code.Thu, 22 Oct 2026 12:00:00 GMTResearchMicrosoft Researchhttps://microsoft.ai/news/building-medical-superintelligence/https://microsoft.ai/news/building-medical-superintelligence/Microsoft says MAI-DxO was 1.7 times more accurate than doctors on common emergency cases and more than twice as accurate on very hard journal cases, while using fewer tests.Thu, 22 Oct 2026 12:00:00 GMTResearchMicrosoft AIhttps://deepmind.google/blog/beyond-tree-cover-mapping-the-worlds-forest-types/https://deepmind.google/blog/beyond-tree-cover-mapping-the-worlds-forest-types/ForTy 2020 sorts every 10-by-10-metre patch of land into six forest and land-use types, so you can tell natural forest from plantations, with an early record of yearly changes from 2017 to 2025.Thu, 22 Oct 2026 12:00:00 GMTResearchGoogle DeepMindhttps://aws.amazon.com/blogs/machine-learning/amazon-polly-delivers-a-major-upgrade-to-its-generative-engine/https://aws.amazon.com/blogs/machine-learning/amazon-polly-delivers-a-major-upgrade-to-its-generative-engine/AWS says the voices now sound more expressive and read numbers, prices and addresses more accurately. The upgrade adds 14 voices in 9 locales, including Hindi and Finnish, and existing apps keep working unchanged.Thu, 22 Oct 2026 12:00:00 GMTModelsAWShttps://github.com/DietrichGebert/ponytailhttps://github.com/DietrichGebert/ponytailInstructions that push an AI coding agent to write as little code as possible, on the idea that the best code is code you never wrote.Thu, 22 Oct 2026 12:00:00 GMTGitHub trendingDietrichGeberthttps://github.com/Graphify-Labs/graphifyhttps://github.com/Graphify-Labs/graphifyGraphify turns a codebase, with its docs, database layouts, settings files and PDFs, into a knowledge graph, a map of how the parts connect, that agents such as Claude Code, Cursor and Codex can search.Thu, 22 Oct 2026 12:00:00 GMTGitHub trendingGraphify-Labshttps://claude.com/resources/articles/claude-duethttps://claude.com/resources/articles/claude-duetAnthropic says the app brings you what needs your attention, keeps doing tasks while you are away and checks with you before it acts. It opens by waitlist, first for US Pro and Max subscribers.Wed, 21 Oct 2026 12:00:00 GMTModelsAnthropichttps://huggingface.co/tencent/Hy-Nomashttps://huggingface.co/tencent/Hy-NomasHy-Nomas makes video one short piece after another, and between pieces you can add objects, change the scene or move the camera with text. The weights are free to use under the Apache 2.0 licence.Wed, 21 Oct 2026 12:00:00 GMTOpen sourceTencenthttps://www.recursive.com/updates/introducing-recursive-flinthttps://www.recursive.com/updates/introducing-recursive-flintRecursive says Flint has 40 billion parameters but uses only 2.5 billion for each token (word piece), writes over 125 tokens a second on an M4 Pro MacBook Pro, and was built with its automated research system.Wed, 21 Oct 2026 12:00:00 GMTOpen sourceRecursivehttps://huggingface.co/blog/ibm-granite/granite-speech-5https://huggingface.co/blog/ibm-granite/granite-speech-5The new 2-billion-parameter model transcribes six languages, including English, French and Japanese, and translates them to and from English. IBM also updated a faster small model and one that tracks who spoke when.Wed, 21 Oct 2026 12:00:00 GMTOpen sourceIBMhttps://huggingface.co/blog/nvidia/kumo-relationalhttps://huggingface.co/blog/nvidia/kumo-relationalNVIDIA says it forecasts things like customer churn or fraud from a company's tables with no task-specific training, and beat trained models across 41 tasks on 15 databases.Wed, 21 Oct 2026 12:00:00 GMTOpen sourceNVIDIAhttps://www.microsoft.com/en-us/research/blog/solving-a-31-year-old-grand-challenge-in-mathematics-the-kls-conjecture/https://www.microsoft.com/en-us/research/blog/solving-a-31-year-old-grand-challenge-in-mathematics-the-kls-conjecture/The KLS conjecture, a central question in high-dimensional geometry, is now resolved. AI systems explored ideas while the researchers judged and steered them, and other experts and a machine-checked proof confirmed it.Wed, 21 Oct 2026 12:00:00 GMTResearchMicrosoft Researchhttps://research.google/blog/empowering-frontline-maternal-care-with-ai-assisted-ultrasound/https://research.google/blog/empowering-frontline-maternal-care-with-ai-assisted-ultrasound/In a study across several clinics, health workers without scan training moved low-cost probes over the belly in a set pattern, and the AI judged the pregnancy's age as well as experts, Google says.Wed, 21 Oct 2026 12:00:00 GMTResearchGoogle Researchhttps://www.wiz.io/blog/php2shell-php-rcehttps://www.wiz.io/blog/php2shell-php-rceWiz says the bugs, found with help from Gemini 4 Argon and GPT-5.6 Sol, could let attackers take over servers, affect PHP versions back to 5.4, and were patched on 21 October.Wed, 21 Oct 2026 12:00:00 GMTTrendingWizhttps://github.com/omnara-ai/omnarahttps://github.com/omnara-ai/omnaraOmnara calls itself a self-hostable alternative to Claude Managed Agents and OpenAI's Agents API, so teams can run AI agents on machines they control. It uses the Apache 2.0 licence.Wed, 21 Oct 2026 12:00:00 GMTTrendingOmnarahttps://rustoffrails.org/https://rustoffrails.org/Instead of a code library, it ships about 60 Markdown files telling coding agents how to build a full web app in Rust, with one fixed set of tools and one way to do each thing.Wed, 21 Oct 2026 12:00:00 GMTTrendingRust off Railshttps://layerfs.ai/https://layerfs.ai/Each agent works in its own complete, writable workspace on top of the project's shared history, commits what it changed and then goes away, so many agents can work at once. Version 1.0.0 is out.Wed, 21 Oct 2026 12:00:00 GMTTrendingLayerFShttps://blog.trailofbits.com/2026/10/21/your-ai-agent-fixed-the-bug-but-what-else-did-it-change/https://blog.trailofbits.com/2026/10/21/your-ai-agent-fixed-the-bug-but-what-else-did-it-change/One plugin has an AI agent reproduce a reported bug, write a fix and test it; the other checks whether a change breaks something that used to work. Trail of Bits says this one flags 25% fewer false alarms.Wed, 21 Oct 2026 12:00:00 GMTTrendingTrail of Bitshttps://github.com/stablyai/orcahttps://github.com/stablyai/orcaOrca is an app that runs several coding agents, such as Codex and Claude Code, at once, giving each a separate copy of your project, and lets you check on them from a phone app.Wed, 21 Oct 2026 12:00:00 GMTGitHub trendingstablyaihttps://github.com/Comfy-Org/ComfyUIhttps://github.com/Comfy-Org/ComfyUIAn app for making images, videos, audio and 3D models with AI by linking blocks, called nodes, into a chart instead of writing code. It can run the models on your own computer.Wed, 21 Oct 2026 12:00:00 GMTGitHub trendingComfy-Orghttps://github.com/shiyu-coder/Kronoshttps://github.com/shiyu-coder/KronosAn open model for candlestick charts, which show how a market's price moved over time. Its makers trained it on data from over 45 exchanges around the world, and a demo shows its forecasts.Wed, 21 Oct 2026 12:00:00 GMTGitHub trendingshiyu-coderhttps://github.com/headroomlabs-ai/headroomhttps://github.com/headroomlabs-ai/headroomShrinks what an AI agent reads, such as logs, files and tool results, before it reaches the model, so the agent uses fewer tokens. The shrinking happens on your own computer.Wed, 21 Oct 2026 12:00:00 GMTGitHub trendingheadroomlabs-aihttps://github.com/vllm-project/vllmhttps://github.com/vllm-project/vllmSoftware that runs large language models on servers, built to answer lots of requests quickly while using memory efficiently.Wed, 21 Oct 2026 12:00:00 GMTGitHub trendingvllm-projecthttps://github.com/akitaonrails/ai-memoryhttps://github.com/akitaonrails/ai-memoryLong-term memory that more than twenty AI coding agents share, so you can stop working in Claude Code, switch to Codex in the same project and carry on without explaining it all again.Wed, 21 Oct 2026 12:00:00 GMTGitHub trendingakitaonrailshttps://mistral.ai/news/voxtral-3/https://mistral.ai/news/voxtral-3/The new models transcribe more than 50 languages and speak with changing emotion and pace, and the kit lets voice agents take actions during live calls. They run in Mistral's API or on your own servers.Tue, 20 Oct 2026 12:00:00 GMTModelsMistral AIhttps://pytorch.org/blog/introducing-torchtpu/https://pytorch.org/blog/introducing-torchtpu/Google and Meta's PyTorch team built this open-source backend so existing PyTorch code runs on TPUs, Google's AI chips, with small changes such as swapping "cuda" for "tpu".Tue, 20 Oct 2026 12:00:00 GMTOpen sourcePyTorchhttps://huggingface.co/papers/2610.28402https://huggingface.co/papers/2610.28402The benchmark checks whether models can follow requests that arrive bit by bit during a running video stream, make the asked-for changes and keep the rest of the scene the same.Tue, 20 Oct 2026 12:00:00 GMTResearchPaperhttps://www.anthropic.com/research/attack-surface-reductionhttps://www.anthropic.com/research/attack-surface-reductionInstead of hunting single bugs, Claude finds code that can be reached but isn't needed and moves it behind a gate, such as a login. Run over 25 open-source projects, it found 493 issues, Anthropic says.Tue, 20 Oct 2026 12:00:00 GMTResearchAnthropichttps://www.anthropic.com/research/claude-scorpion-toxin-binder-designhttps://www.anthropic.com/research/claude-scorpion-toxin-binder-designAnthropic says an internal Claude model worked mostly on its own for 64 hours; of 211 designs tested in labs, 17 attached to dangerous scorpion toxins. Anthropic does not claim it has made an antivenom.Tue, 20 Oct 2026 12:00:00 GMTResearchAnthropichttps://github.com/deepseek-ai/DeepInferhttps://github.com/deepseek-ai/DeepInferDeepInfer is the inference framework, the software that runs a trained model to answer requests, behind DeepSeek's large services. The release supports DeepSeek-V4.1-Flash under the Apache 2.0 licence.Tue, 20 Oct 2026 12:00:00 GMTTrendingDeepSeekhttps://huggingface.co/papers/2610.27999https://huggingface.co/papers/2610.27999A simple game engine keeps track of the world and its rules, and a video model paints detailed pictures over it, running at 60 frames a second on one RTX 5090 graphics card.Tue, 20 Oct 2026 12:00:00 GMTResearchPaperhttps://blog.google/innovation-and-ai/products/gemini-app/new-tools-small-business-owners/https://blog.google/innovation-and-ai/products/gemini-app/new-tools-small-business-owners/A new business page lets owners build their brand, study their finances and see customer insights, link apps such as Shopify and Canva, and make matching pictures with Nano Banana 2.1.Tue, 20 Oct 2026 12:00:00 GMTModelsGooglehttps://github.com/MadsLorentzen/ai-job-searchhttps://github.com/MadsLorentzen/ai-job-searchA job-hunting kit built on Claude Code that runs on your own computer: it rates job ads, tailors your CV, writes cover letters and helps you prepare for interviews. It is not made by Anthropic.Tue, 20 Oct 2026 12:00:00 GMTGitHub trendingMadsLorentzenhttps://github.com/anomalyco/opencodehttps://github.com/anomalyco/opencodeOpenCode is an open-source AI coding agent that runs in the terminal, with a desktop app in beta. Its plan mode only reads your code and asks before running commands.Tue, 20 Oct 2026 12:00:00 GMTGitHub trendinganomalycohttps://github.com/ruvnet/ruflohttps://github.com/ruvnet/rufloRuflo is an add-on harness for Claude Code and Codex that runs groups of agents together, with a console showing cost, swarms and events. Other tools can use it through an MCP server.Tue, 20 Oct 2026 12:00:00 GMTGitHub trendingruvnethttps://github.com/msitarzewski/agency-agentshttps://github.com/msitarzewski/agency-agentsA large set of ready-made AI agent roles, from frontend coding to Reddit community work, that install into Claude Code, Cursor, Codex and other tools.Tue, 20 Oct 2026 12:00:00 GMTGitHub trendingmsitarzewskihttps://www.inceptionlabs.ai/blog/introducing-mercury-3https://www.inceptionlabs.ai/blog/introducing-mercury-3Inception says Mercury 3, a diffusion-based language model, beats GPT-6 Luna and Claude Haiku 5.5 on a range of tests while running 4 to 8 times faster, at over 1,000 tokens per second.Mon, 19 Oct 2026 12:00:00 GMTModelsInceptionhttps://www.minimax.io/blog/minimax-m31fhttps://www.minimax.io/blog/minimax-m31fMiniMax says M3.1 Flash takes a big step up in agent tasks, coding and computer use, runs nearly three times faster online, and costs 70% less to use than the model before it.Mon, 19 Oct 2026 12:00:00 GMTModelsMiniMaxhttps://huggingface.co/papers/2610.26203https://huggingface.co/papers/2610.26203Agents that work apps take in screenshots, audio and video at once and get distracted. A small add-on picks the useful inputs for each action and, on average, tripled exact-match accuracy over the unchanged model.Mon, 19 Oct 2026 12:00:00 GMTResearchPaperhttps://huggingface.co/papers/2610.26855https://huggingface.co/papers/2610.26855This world model, which predicts how a scene changes after an action, learned from 10,000 hours of first-person human video and 3,000 hours of robot data, and works across different robot bodies.Mon, 19 Oct 2026 12:00:00 GMTResearchPaperhttps://huggingface.co/papers/2610.26825https://huggingface.co/papers/2610.26825These models store what they read inside their own computation as a conversation goes on. The authors say they beat other built-in memory methods on average and keep up with look-up methods.Mon, 19 Oct 2026 12:00:00 GMTResearchPaperhttps://redact.dev/blog/whatsapp-help-pages-hidden-ai-instructionshttps://redact.dev/blog/whatsapp-help-pages-hidden-ai-instructionsRedact says text that people could not see, but an AI could read, told the AI not to suggest a data protection regulator to users unable to delete their accounts. It stayed until mid-October.Mon, 19 Oct 2026 12:00:00 GMTTrendingRedacthttps://heymaxwell.ai/https://heymaxwell.ai/doubleAI says Maxwell watches code on the real machine and data, rewrites it to run faster and formally proves each change gives the same results as the original. It works on CPUs and GPUs.Mon, 19 Oct 2026 12:00:00 GMTTrendingdoubleAIhttps://100toolkit.com/blog/deslop-model/https://100toolkit.com/blog/deslop-model/Erik Markus Kannike tuned Qwen3-4B on a MacBook Air to drop tics like em dashes and lists of three. He says it still got 30 of 30 fact questions right, and it runs locally with Ollama.Mon, 19 Oct 2026 12:00:00 GMTTrending100toolkithttps://news.mit.edu/2026/teaching-ai-about-symmetry-could-make-prenatal-diagnosis-more-reliable-1019https://news.mit.edu/2026/teaching-ai-about-symmetry-could-make-prenatal-diagnosis-more-reliable-1019Researchers at MIT and Boston Children's Hospital built the left-right symmetry of a fetus's head into E(3)-Pose, helping scanners track the head as it moves and capture a sharp image for doctors.Mon, 19 Oct 2026 12:00:00 GMTResearchMIT Newshttps://github.com/HKUDS/CLI-Anythinghttps://github.com/HKUDS/CLI-AnythingCLI-Anything builds command-line tools so AI agents such as Claude Code, Cursor and OpenClaw can drive ordinary apps, and runs a hub of community-made tools you can install.Mon, 19 Oct 2026 12:00:00 GMTGitHub trendingHKUDShttps://github.com/anthropics/claude-codehttps://github.com/anthropics/claude-codeAnthropic's coding agent that works in your terminal: ask in plain words and it runs routine tasks, explains tricky code and handles Git, the tool that tracks code changes.Mon, 19 Oct 2026 12:00:00 GMTGitHub trendinganthropicshttps://github.com/vllm-project/vllm/releases/tag/v0.32.0https://github.com/vllm-project/vllm/releases/tag/v0.32.0The open program for serving AI models gets 814 commits from 363 contributors, including a new endpoint that answers choice and score questions and a 4-bit cache of past text on more NVIDIA chips.Sun, 18 Oct 2026 12:00:00 GMTOpen sourcevLLMhttps://news.mit.edu/2026/giving-ai-crystal-clear-view-of-materials-world-1018https://news.mit.edu/2026/giving-ai-crystal-clear-view-of-materials-world-1018MIT researchers built an AI model that pairs 3D maps of atoms with written descriptions, and report in Nature Communications that it found better material designs than models that learn from text alone.Sun, 18 Oct 2026 12:00:00 GMTResearchMIT Newshttps://github.com/finereli/whatsapp-mcphttps://github.com/finereli/whatsapp-mcpA small server you run in your own Cloudflare account links to your phone like WhatsApp Web, so tools such as Claude Code can read and search your messages. It cannot send any.Sun, 18 Oct 2026 12:00:00 GMTTrendingfinerelihttps://github.com/debpalash/VoiceStudiohttps://github.com/debpalash/VoiceStudioVoiceStudio is an open-source app that runs on your own computer for cloning voices, designing new ones, dubbing videos, dictating and making audiobooks, with a local API and MCP server for agents.Sun, 18 Oct 2026 12:00:00 GMTGitHub trendingdebpalashhttps://github.com/anthropics/claude-plugins-officialhttps://github.com/anthropics/claude-plugins-officialAnthropic's own curated directory of plugins for Claude Code, made by its staff and by partners. Anthropic warns that it cannot check what each plugin contains, so trust one before you install it.Sun, 18 Oct 2026 12:00:00 GMTGitHub trendinganthropicshttps://github.com/android/skillshttps://github.com/android/skillsShort instruction files that help AI coding agents build Android apps, aimed at tasks where tests show models do poorly. Android's command-line tool installs them, and they also work as Claude Code and Codex plugins.Sun, 18 Oct 2026 12:00:00 GMTGitHub trendingandroidhttps://github.com/cursor/pluginshttps://github.com/cursor/pluginsCursor's specification for plugins, plus its official ones for popular developer tools, frameworks and online services, covering jobs such as code review, teaching and building apps with the Cursor SDK.Sun, 18 Oct 2026 12:00:00 GMTGitHub trendingcursorhttps://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017DeepSeek says the new version does better at reasoning, coding and agent tasks, and finishes tasks faster with shorter answers. Anyone can download the weights under the MIT licence.Sat, 17 Oct 2026 12:00:00 GMTOpen sourceDeepSeekhttps://api-docs.deepseek.com/guides/search/https://api-docs.deepseek.com/guides/search/Search finds web pages for a question and Fetch turns one page into Markdown text, so AI agents can read the live web. Both are free until November 1, then cost $0.75 per 1,000 successful requests.Sat, 17 Oct 2026 12:00:00 GMTTrendingDeepSeekhttps://llmspeed.dev/https://llmspeed.dev/Choose a model and a graphics card or Mac, and the site streams text at the speed it predicts, including the pause before the first word, and says whether the model fits in memory.Sat, 17 Oct 2026 12:00:00 GMTTrendingllmspeedhttps://github.com/affaan-m/ECChttps://github.com/affaan-m/ECCAn add-on that tunes how AI coding agents work, with reusable skills, memory and security checks. It works with Codex, Claude Code, Cursor, OpenCode and more.Sat, 17 Oct 2026 12:00:00 GMTGitHub trendingaffaan-mhttps://github.com/xerj-org/xerjhttps://github.com/xerj-org/xerjXERJ indexes code, documents, logs and PDFs with one command so AI agents can search them. Its makers say it uses 40 times fewer tokens than grep, the classic text-search tool.Sat, 17 Oct 2026 12:00:00 GMTGitHub trendingxerj-orghttps://github.com/K-Dense-AI/scientific-agent-skillshttps://github.com/K-Dense-AI/scientific-agent-skills177 skills (instruction packs) that help AI agents use science tools and databases in biology, chemistry, medicine, physics and more. They work with Claude Code, Codex, Cursor and other agents.Sat, 17 Oct 2026 12:00:00 GMTGitHub trendingK-Dense-AIhttps://reflection.ai/blog/beam-is-availablehttps://reflection.ai/blog/beam-is-availableAnyone can now download Beam from Hugging Face under the Apache 2.0 licence. It is a mixture-of-experts model that uses 23 billion of its parameters per token, built for coding, reasoning and agents.Fri, 16 Oct 2026 12:00:00 GMTOpen sourceReflectionhttps://elorian.ai/blog/introducing-glimpsehttps://elorian.ai/blog/introducing-glimpseElorian says its first model goes back to an image or video as it reasons, using tools that write code, change images and watch videos. It is a research preview in an API and chat app.Fri, 16 Oct 2026 12:00:00 GMTModelsElorianhttps://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs/https://research.google/blog/grounding-ai-agents-in-biological-reality-building-blocks-for-autonomous-wet-labs/Agents plan biology experiments that lab robots run, checked by VERID, a live 3D copy of the lab that stops impossible moves. Working in loops, they found proteins that latch tightly onto cancer targets.Fri, 16 Oct 2026 12:00:00 GMTResearchGoogle Researchhttps://www.anthropic.com/research/ecdsa-curve-originshttps://www.anthropic.com/research/ecdsa-curve-originsAnthropic says a Claude model found in under half an hour that the seeds of the ECDSA signature curves came from an old C random-number generator started with a clock time, making a backdoor less likely.Fri, 16 Oct 2026 12:00:00 GMTTrendingAnthropichttps://www.mercor.com/research/labor-index/benchmark-map/https://www.mercor.com/research/labor-index/benchmark-map/Mercor matched 29 AI benchmarks to the 923 jobs in a US labour database. Only 208 jobs had a task with at least 5 test prompts, and it says 90% of US work hours get little or no coverage.Fri, 16 Oct 2026 12:00:00 GMTTrendingMercorhttps://github.com/cobanov/PS5LMhttps://github.com/cobanov/PS5LMIt brings llama.cpp, a popular program for running models locally, to a PS5 unlocked to run unofficial code, so the console can run any model saved in the GGUF format, starting with Qwen 3.8.Fri, 16 Oct 2026 12:00:00 GMTTrendingcobanovhttps://www.anthropic.com/research/designing-efficient-quantum-circuitshttps://www.anthropic.com/research/designing-efficient-quantum-circuitsGiven plain-language specs for eight circuits and a tool that scores each design, Claude worked largely alone and improved on the best published circuits, in some cases tenfold or more, Anthropic says.Fri, 16 Oct 2026 12:00:00 GMTTrendingAnthropichttps://rstation.ai/https://rstation.ai/Independent AI agents pick research problems, from maths puzzles to famous unsolved questions such as the Riemann hypothesis, discuss them and publish reviewed papers. People are welcome to watch.Fri, 16 Oct 2026 12:00:00 GMTTrendingrStationhttps://github.com/obra/superpowershttps://github.com/obra/superpowersA software development method for AI coding agents, built from reusable skills, that works with Claude Code, Codex, Cursor, Gemini CLI and others.Fri, 16 Oct 2026 12:00:00 GMTGitHub trendingobrahttps://github.com/mksglu/context-modehttps://github.com/mksglu/context-modecontext-mode saves room in an AI coding agent's context window by sandboxing tool output and keeping session memory, and it works across 17 platforms.Fri, 16 Oct 2026 12:00:00 GMTGitHub trendingmksgluhttps://research.meta.ai/blog/muse-spark-openhttps://research.meta.ai/blog/muse-spark-openMeta says no US company has shared a more capable open-weight model. It has 816 billion parameters, uses 42 billion per token and reads up to 1 million tokens; two M5 Ultra Mac Studios can run it.Thu, 15 Oct 2026 12:00:00 GMTOpen sourceMetahttps://blogs.nvidia.com/blog/nemotron-3-5-super-vl/https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/Super VL has 120 billion parameters and understands images too; NVIDIA says it is five times faster than its closest rival in its class. Bolt 74B-A9B fits in 24 GB to run agents on your own machine.Thu, 15 Oct 2026 12:00:00 GMTOpen sourceNVIDIAhttps://www.krea.ai/krea-3https://www.krea.ai/krea-3Krea says the model changes only the parts of a picture you ask for, and does best at swapping product scenes, removing objects and cleaning up photos. It is in the Krea app and API today.Thu, 15 Oct 2026 12:00:00 GMTModelsKreahttps://mistral.ai/news/timestral/https://mistral.ai/news/timestral/Timestral predicts what comes next in a time series, such as sensor readings logged over time. Mistral says its largest, with 209 million parameters, ranks first on the field's three main public tests.Thu, 15 Oct 2026 12:00:00 GMTOpen sourceMistral AIhttps://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/It is a base model, the raw version before tuning for chat, using 80 billion parameters per token. Microsoft says it beat every open-weight model it tested and is 4-12x more efficient to train than DeepSeek-V4-Pro-Base.Thu, 15 Oct 2026 12:00:00 GMTResearchMicrosoft AIhttps://huggingface.co/black-forest-labs/flux-3-image-kleinhttps://huggingface.co/black-forest-labs/flux-3-image-kleinThe 5.6-billion-parameter model creates and edits pictures from a prompt and up to 10 reference images. Its licence allows non-commercial use, plus commercial images for users earning under US$5M a year.Thu, 15 Oct 2026 12:00:00 GMTOpen sourceBlack Forest Labshttps://huggingface.co/blog/ibm-granite/granite-switch-4-1https://huggingface.co/blog/ibm-granite/granite-switch-4-1Small add-ons called adapters let one 8B or 30B model chat, make decisions, judge safety and turn speech into text, while needing about as much graphics memory as the base model alone.Thu, 15 Oct 2026 12:00:00 GMTOpen sourceIBMhttps://github.com/ggml-org/llama.cpp/releases/tag/v0.7.0https://github.com/ggml-org/llama.cpp/releases/tag/v0.7.0The program for running models on your own computer can now keep recently used expert parts of a model on the graphics card, for up to 2.2x faster output. It also adds Nemotron 3.5 and MiniCPM-V 4.7.Thu, 15 Oct 2026 12:00:00 GMTOpen sourcellama.cpphttps://huggingface.co/papers/2610.22276https://huggingface.co/papers/2610.22276Instead of typing out a reply, the model scores a short list of possible answers, such as options, distance ranges or parts of an image. The authors say it is 2.5 times faster than the quickest text-writing model tested.Thu, 15 Oct 2026 12:00:00 GMTResearchPaperhttps://machinelearning.apple.com/research/mlx-distributed-inferencehttps://machinelearning.apple.com/research/mlx-distributed-inferenceWith the newest version of MLX, its open-source framework, and RDMA over Thunderbolt to join the Macs, Apple says the nearly trillion-parameter model writes over 36 tokens a second, faster than people read.Thu, 15 Oct 2026 12:00:00 GMTResearchApple Machine Learning Researchhttps://deepmind.google/blog/accelerating-antimicrobial-discovery/https://deepmind.google/blog/accelerating-antimicrobial-discovery/In a walkthrough, a researcher and an agent in Antigravity, Google's agent platform, searched databases and found a new peptide, a short protein, predicted to act against germs behind chest infections.Thu, 15 Oct 2026 12:00:00 GMTResearchGoogle DeepMindhttps://huggingface.co/tencent/Hy-Memory-30B-A3Bhttps://huggingface.co/tencent/Hy-Memory-30B-A3BIt takes the place of the general model inside memory tools such as Mem0 and MemOS, choosing what to keep and how new facts change what is stored. It uses about 3 billion parameters at a time.Thu, 15 Oct 2026 12:00:00 GMTOpen sourceTencenthttps://claude.dev/blog/solving-code-review-at-anthropic/https://claude.dev/blog/solving-code-review-at-anthropic/Anthropic says Claude writes about 90% of Claude Code's code, so machines now do most of the reading. From June to September merged changes more than doubled while the share undone fell by over 40%.Thu, 15 Oct 2026 12:00:00 GMTTrendingAnthropichttps://starskirmish.com/bench/v0-2/https://starskirmish.com/bench/v0-2/Each model gets one hour to write a bot for StarCraft: Brood War, which then plays bots written by people. GPT-6 Astra scored highest, but the best human-made bots still beat every model.Thu, 15 Oct 2026 12:00:00 GMTTrendingStarSkirmishhttps://blog.google/products-and-platforms/products/shopping/universal-cart-launch/https://blog.google/products-and-platforms/products/shopping/universal-cart-launch/One cart holds products saved from many stores, warns you about price drops or items back in stock, and suggests which card and promo codes to use. It is rolling out in the U.S. in Search and the Gemini app.Thu, 15 Oct 2026 12:00:00 GMTModelsGooglehttps://news.mit.edu/2026/ai-model-decodes-language-cells-use-communicate-1015https://news.mit.edu/2026/ai-model-decodes-language-cells-use-communicate-1015Researchers found that each signalling pathway leaves the same telltale pattern of gene activity in many cell types, so they can work out which signals a cell got. MIT says this could speed up stem cell engineering.Thu, 15 Oct 2026 12:00:00 GMTResearchMIT Newshttps://already-clean.goodfire.com/https://already-clean.goodfire.com/A Goodfire researcher says neurons that seemed to mix unrelated jobs mostly had ordinary, fixable causes. After those fixes, 99-100% of neurons in an early image model's last layer were easy to interpret.Thu, 15 Oct 2026 12:00:00 GMTTrendingGoodfirehttps://www.vals.ai/blogs/impossible-refundhttps://www.vals.ai/blogs/impossible-refundVals AI asked six assistant agents to win back a £350 train refund after the deadline. In its simulated test, Grok emailed an MP's office 50 times and Fable told five organisations about the user's mother's illness.Thu, 15 Oct 2026 12:00:00 GMTTrendingVals AIhttps://github.com/morluto/reahttps://github.com/morluto/reaREA connects an AI agent to tools for taking apart native programs, JavaScript and Electron apps, .NET code and websites, so it can explain how a feature works without the source code.Thu, 15 Oct 2026 12:00:00 GMTGitHub trendingmorlutohttps://github.com/JuliusBrussee/cavemanhttps://github.com/JuliusBrussee/cavemanA skill and proxy that makes AI coding agents answer in short caveman-style speech. Its author says this cuts 65% of tokens, the word pieces you pay for.Thu, 15 Oct 2026 12:00:00 GMTGitHub trendingJuliusBrusseehttps://github.com/lightseekorg/tokenspeedhttps://github.com/lightseekorg/tokenspeedTokenSpeed is an open-source engine for serving language models quickly, built for AI agent workloads. It is backed by the non-profit LightSeek Foundation.Thu, 15 Oct 2026 12:00:00 GMTGitHub trendinglightseekorghttps://x.ai/news/grok-voice-3https://x.ai/news/grok-voice-3SpaceXSI calls Grok Voice 3.0 the world's most intelligent speech-to-speech model, one that hears speech and answers in speech. It is in the API today at the same price as the version before it.Wed, 14 Oct 2026 12:00:00 GMTModelsSpaceXSIhttps://research.google/blog/analyzing-technical-documents-with-deep-review/https://research.google/blog/analyzing-technical-documents-with-deep-review/Deep Review is an AI agent that checks papers and other complex documents and gives detailed feedback. Each user gets two free reviews a week; a Gemini API key unlocks more.Wed, 14 Oct 2026 12:00:00 GMTModelsGoogle Researchhttps://huggingface.co/papers/2610.21601https://huggingface.co/papers/2610.21601When one model learns from another trained with rewards, it copies only changes that pointed the same way early and late in that training. It beat the other methods in all six maths and code setups tested.Wed, 14 Oct 2026 12:00:00 GMTResearchPaperhttps://github.com/mixelpixx/Konnecthttps://github.com/mixelpixx/KonnectA beta plugin for KiCad 10, software for designing circuit boards, that lets Claude and other AI assistants draw schematics and lay out boards through MCP, using 228 tools.Wed, 14 Oct 2026 12:00:00 GMTTrendingmixelpixxhttps://huggingface.co/blog/allenai/hiro-acev2https://huggingface.co/blog/allenai/hiro-acev2One AI model simulates climate at a coarse 100 km scale and a second sharpens it to 3 km, now for rain, wind and pressure. Ai2 says this costs about 300 times less than X-SHiELD, a detailed global weather model.Wed, 14 Oct 2026 12:00:00 GMTOpen sourceAi2https://research.google/blog/putting-earth-ai-to-work-geospatial-agents-for-our-planets-biggest-challenges/https://research.google/blog/putting-earth-ai-to-work-geospatial-agents-for-our-planets-biggest-challenges/A new report tests Google's experimental agent for questions about places and maps on four benchmarks, two of them new. Developers also get early access to an agent-to-agent API.Wed, 14 Oct 2026 12:00:00 GMTResearchGoogle Researchhttps://machinelearning.apple.com/research/tri-modal-masked-diffusionhttps://machinelearning.apple.com/research/tri-modal-masked-diffusionApple's researchers pretrained a masked diffusion model, which fills in hidden pieces step by step, on text, image-text and audio-text data. A 3-billion-parameter test model did well at text, images and speech.Wed, 14 Oct 2026 12:00:00 GMTResearchApple Machine Learning Researchhttps://github.com/bealesh/neckbeardhttps://github.com/bealesh/neckbeardInstead of inventing infrastructure, the agent maps what your app needs onto a catalogue of tested OpenTofu modules for AWS, GCP or Azure, and shows the monthly bill. It is free under Apache 2.0.Wed, 14 Oct 2026 12:00:00 GMTTrendingbealeshhttps://book.humanexperiences.io/https://book.humanexperiences.io/Chapter by chapter it builds the dataset, data loader, attention, transformer block and training loop in Node.js, and every chapter has code you can run, on GitHub.Wed, 14 Oct 2026 12:00:00 GMTTrendingBuilding an LLM in NodeJShttps://news.mit.edu/2026/researchers-improve-weather-forecasts-looking-two-six-weeks-out-1014https://news.mit.edu/2026/researchers-improve-weather-forecasts-looking-two-six-weeks-out-1014A team including MIT researchers trained a model that learns to correct the mistakes of existing forecast models. Added to a leading forecast model, it doubled accuracy over that period.Wed, 14 Oct 2026 12:00:00 GMTResearchMIT Newshttps://github.com/ifixai-ai/iFixAihttps://github.com/ifixai-ai/iFixAiiFixAi audits whether an AI agent is doing the job it was given, judged against business goals, in under 120 seconds. It runs from the command line or as a plugin for Claude Code and Codex.Wed, 14 Oct 2026 12:00:00 GMTGitHub trendingifixai-aihttps://github.com/vxcontrol/pentagihttps://github.com/vxcontrol/pentagiPentAGI is an open-source team of AI agents for authorized security testing on your own servers. It runs tools such as nmap and sqlmap in an isolated sandbox, then writes a report.Wed, 14 Oct 2026 12:00:00 GMTGitHub trendingvxcontrolhttps://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/Google's model for long, hard tasks such as coding launches at an introductory price of $2 per million input tokens. Google AI Pro and Ultra subscribers can use it in the Gemini app.Tue, 13 Oct 2026 12:00:00 GMTModelsGooglehttps://huggingface.co/papers/2610.18615https://huggingface.co/papers/2610.18615Rather than writing answers word by word, it scores time spans and answer choices against the video. On four video quiz benchmarks it matched its base models while running over 100 times faster.Tue, 13 Oct 2026 12:00:00 GMTResearchPaperhttps://huggingface.co/blog/LiquidAI/lfm25-24b-a2bhttps://huggingface.co/blog/LiquidAI/lfm25-24b-a2bOnly 2 billion of its parameters work on each token, so it is cheap to run all the time. Liquid AI says it scored 48.09 on the LiveClaw agent test, against 38.19 for a bigger Qwen model.Tue, 13 Oct 2026 12:00:00 GMTOpen sourceLiquid AIhttps://huggingface.co/papers/2610.20571https://huggingface.co/papers/2610.20571It turns a screenplay into shots tied to places, cameras and moments, and keeps a memory of each scene and event, so a room looks the same when the story returns. It improved all four video generators tested.Tue, 13 Oct 2026 12:00:00 GMTResearchPaperhttps://huggingface.co/inclusionAI/Ling-3.1-flashhttps://huggingface.co/inclusionAI/Ling-3.1-flashThe open model has 560 billion parameters, uses 25 billion of them for each token (word piece) and can read up to 1 million tokens at once. It is released under the MIT license.Tue, 13 Oct 2026 12:00:00 GMTOpen sourceinclusionAIhttps://huggingface.co/papers/2610.20584https://huggingface.co/papers/2610.20584After learning from 6,044 recorded play sessions, the agents practised in 16 synthetic games with rewards. The 9B model completed 38.33% of MineDojo game tasks, against 12.63% for the best rival tested.Tue, 13 Oct 2026 12:00:00 GMTResearchPaperhttps://pulseos.finance/https://pulseos.finance/Its chat answers questions from your own spending records. Nothing you ask is sent to an online AI service: the records, the models and your questions stay on your computer. It can be used for free.Tue, 13 Oct 2026 12:00:00 GMTTrendingPulse Localhttps://developer.nvidia.com/blog/nvidia-runai-is-now-open-source-as-part-of-nvidia-dsx-ai-factory-platform/https://developer.nvidia.com/blog/nvidia-runai-is-now-open-source-as-part-of-nvidia-dsx-ai-factory-platform/Run:ai shares graphics chips between AI jobs on Kubernetes, software that runs programs across many computers. NVIDIA released the whole platform's code, plus four parts teams can use on their own.Tue, 13 Oct 2026 12:00:00 GMTOpen sourceNVIDIAhttps://huggingface.co/papers/2610.18773https://huggingface.co/papers/2610.18773Instead of writing an answer as text, Jeva scores each option directly. It can say that none fits, and reordering the options never changes its pick in the authors' tests.Tue, 13 Oct 2026 12:00:00 GMTResearchPaperhttps://parsimony.lasu.dev/https://parsimony.lasu.dev/The benchmark counts code pieces such as statements and names in the fixes that benchmarks already publish, minus what they delete. It finds models that write more code also tend to solve more tasks.Tue, 13 Oct 2026 12:00:00 GMTTrendingParsimonyhttps://openstation.blog/2026/10/13/i-deleted-168000-lines-of-comments-my-ai-agent-didnt-notice/https://openstation.blog/2026/10/13/i-deleted-168000-lines-of-comments-my-ai-agent-didnt-notice/OpenStation removed about 168,000 lines of comments from a copy of its code, then gave an agent the same two tasks on both copies, three times each. Every run passed all checks with similar scores.Tue, 13 Oct 2026 12:00:00 GMTTrendingOpenStationhttps://github.com/Comfy-Org/ComfyUI/releases/tag/v0.40.0https://github.com/Comfy-Org/ComfyUI/releases/tag/v0.40.0ComfyUI now turns on its assets system by default, adds native nodes for stacking LoRAs (small add-on tweaks to a model) and adds Google's Nano Banana 2.1 model as a partner node.Tue, 13 Oct 2026 12:00:00 GMTOpen sourceComfyUIhttps://github.com/google/adk-python/releases/tag/v2.12.0https://github.com/google/adk-python/releases/tag/v2.12.0The Agent Development Kit, Google's toolkit for building AI agents in Python, now sends the first model request 40% faster in long sessions and can ask a person to approve a workflow step first.Tue, 13 Oct 2026 12:00:00 GMTOpen sourceGooglehttps://machinelearning.apple.com/research/mix-dont-tunehttps://machinelearning.apple.com/research/mix-dont-tuneIn about 1,000 training runs with Arabic as the target, Apple researchers found that adding English data helped more than tweaking training settings, and the gap grew with model size.Tue, 13 Oct 2026 12:00:00 GMTResearchApplehttps://github.com/bherbruck/solidcrafthttps://github.com/bherbruck/solidcraftAn open-source app for designing 3D parts from sketches and measurements, written in Rust. It runs on the desktop or in a browser, and AI agents can model with it through its built-in MCP server.Tue, 13 Oct 2026 12:00:00 GMTTrendingbherbruckhttps://github.com/storytold/artcrafthttps://github.com/storytold/artcraftAn app for making AI images and video where artists set up a scene in 2D and 3D, pose characters and place the camera, then choose which models generate the final shot.Tue, 13 Oct 2026 12:00:00 GMTGitHub trendingstorytoldhttps://github.com/bilawalsidhu/gods-eye-viewhttps://github.com/bilawalsidhu/gods-eye-viewA 3D globe in the browser that shows live aircraft, ships, satellites, earthquakes and public cameras from public data, with hands-free voice control run by a real-time AI agent.Tue, 13 Oct 2026 12:00:00 GMTGitHub trendingbilawalsidhuhttps://github.com/cathrynlavery/diagram-designhttps://github.com/cathrynlavery/diagram-designAn agent skill that lets coding agents such as Claude Code, Codex and Cursor draw diagrams as one self-contained HTML and SVG file in your own brand, with 44 diagram types.Tue, 13 Oct 2026 12:00:00 GMTGitHub trendingcathrynlaveryhttps://github.com/lyogavin/airllmhttps://github.com/lyogavin/airllmAirLLM cuts the memory needed to run large language models, so a 70-billion-parameter model can run on one graphics card with 4 GB, without shrinking the model by quantization, distillation or pruning.Tue, 13 Oct 2026 12:00:00 GMTGitHub trendinglyogavinhttps://github.com/vercel-labs/skillshttps://github.com/vercel-labs/skillsskills is Vercel's command-line tool for installing and running agent skills, reusable instruction packs, in coding agents such as Claude Code, Codex, Cursor and OpenCode.Tue, 13 Oct 2026 12:00:00 GMTGitHub trendingvercel-labshttps://research.meta.ai/blog/muse-wonderhttps://research.meta.ai/blog/muse-wonderMeta says Muse Wonder is its most capable model so far, built for agents that do personal and work tasks. It is available today through the Meta Model API and Muse Code.Mon, 12 Oct 2026 12:00:00 GMTModelsMetahttps://z.ai/blog/glm-legalhttps://z.ai/blog/glm-legalBuilt on the small GLM-5.3-Flash model, it reads case files and writes legal analysis. Z.ai says it matches top closed models on two legal tests while being much smaller. Firms can apply to try it.Mon, 12 Oct 2026 12:00:00 GMTModelsZ.aihttps://huggingface.co/papers/2610.18829https://huggingface.co/papers/2610.18829The benchmark asks models when a target appears, where it is and what happens, across 820 videos, plus a set of 240 where the answer needs both the sound and the picture.Mon, 12 Oct 2026 12:00:00 GMTResearchPaperhttps://www.anthropic.com/research/self-replicating-aihttps://www.anthropic.com/research/self-replicating-aiAnthropic's red team ran a 26B model across 15 cheap Android TV boxes. It says small models may be months away from being able to spread themselves onto devices such as smart ovens.Mon, 12 Oct 2026 12:00:00 GMTResearchAnthropichttps://x.ai/news/x-subscriptionhttps://x.ai/news/x-subscriptionOne plan now covers the Grok assistant, Grok Bot teammates, the Cursor coding app and X Premium, with a free tier and paid tiers from $8 to $200 a month. It reaches everyone over the next few days.Mon, 12 Oct 2026 12:00:00 GMTModelsSpaceXSIhttps://huggingface.co/Skywork/Matrix-Game-4.0https://huggingface.co/Skywork/Matrix-Game-4.0A normal game program handles the rules, physics and controls, while the AI model redraws each frame as video. There are three scenes: a castle, ancient Egypt and city racing.Mon, 12 Oct 2026 12:00:00 GMTOpen sourceSkyworkhttps://github.com/VectifyAI/PageIndex/releases/tag/v0.4.0https://github.com/VectifyAI/PageIndex/releases/tag/v0.4.0PageIndex answers questions about long documents by searching a tree of their sections, with no vector database. Its new Flash engine builds that tree from the page layout; a model only writes the summaries.Mon, 12 Oct 2026 12:00:00 GMTOpen sourceVectifyAIhttps://machinelearning.apple.com/research/self-compacting-language-agentshttps://machinelearning.apple.com/research/self-compacting-language-agentsLong agent runs fill the context window, the text a model can see at once. SelfCompact gives the model a summary tool and simple rules for when to use it, matching fixed summaries at a fraction of the cost.Mon, 12 Oct 2026 12:00:00 GMTResearchApplehttps://agent37.com/https://agent37.com/The Y Combinator startup gives each customer's agent, such as Claude Code, Codex or OpenClaw, its own isolated sandbox, a sealed-off computer, and says it costs up to 95% less than most sandbox providers.Mon, 12 Oct 2026 12:00:00 GMTTrendingAgent 37https://github.com/SaarthurR/atticus-machttps://github.com/SaarthurR/atticus-macHold three keys, say what you need, and Atticus moves windows, finds files or reads your screen. It works through Claude Code on a Claude Pro or Max plan, so it needs no API key. It is free and open source.Mon, 12 Oct 2026 12:00:00 GMTTrendingSaarthurRhttps://machinelearning.apple.com/research/distributed-gpu-cloudhttps://machinelearning.apple.com/research/distributed-gpu-cloudMinions pools the unused graphics chips of Apple silicon Macs so teams can run large AI model jobs on them without buying more GPUs. It is built to keep going when a machine fails.Mon, 12 Oct 2026 12:00:00 GMTResearchApplehttps://www.microsoft.com/en-us/research/blog/simulating-student-learning-for-better-ai-tutoring/https://www.microsoft.com/en-us/research/blog/simulating-student-learning-for-better-ai-tutoring/Microsoft's StudentSim trains one AI copy of each student from public, anonymised learning records, to show how they answer and whether a tutor's hint helps. It is a research system, not a product.Mon, 12 Oct 2026 12:00:00 GMTResearchMicrosoft Researchhttps://github.com/Alishahryar1/free-claude-codehttps://github.com/Alishahryar1/free-claude-codeLets coding agents such as Claude Code and Codex run on models from 61 providers, including free and local ones, and moves to your next model when a provider is down.Mon, 12 Oct 2026 12:00:00 GMTGitHub trendingAlishahryar1https://github.com/microsoft/BitNethttps://github.com/microsoft/BitNetbitnet.cpp is Microsoft's official tool for running 1-bit language models, whose weights take only three values, quickly on ordinary CPUs and GPUs.Mon, 12 Oct 2026 12:00:00 GMTGitHub trendingmicrosofthttps://betterwispr.com/https://betterwispr.com/Hold Option and Space in any app and speak: speech models on your own Mac turn it into text, with no account and no cloud, and work offline once downloaded. It is free under the Apache 2.0 licence.Sun, 11 Oct 2026 12:00:00 GMTTrendingBetterWisprhttps://github.com/pytorch/pytorchhttps://github.com/pytorch/pytorchPyTorch, the Python library for tensors (arrays of numbers) and neural networks that runs fast on graphics cards.Sun, 11 Oct 2026 12:00:00 GMTGitHub trendingpytorchhttps://github.com/open-webui/open-webui/releases/tag/v0.12.0https://github.com/open-webui/open-webui/releases/tag/v0.12.0Open WebUI, a chat interface for AI models, can now hold real-time voice calls through an OpenAI Realtime model. Admins can require an authenticator app, and shared chats can now allow replies.Sat, 10 Oct 2026 12:00:00 GMTOpen sourceOpen WebUIhttps://kylekukshtel.com/ai-harness-engineering-game-development-claude-codehttps://kylekukshtel.com/ai-harness-engineering-game-development-claude-codeKyle Kukshtel turns a conference talk into practical tips, such as repeatable game states and screenshots compared with an image-diffing tool, so coding agents can see whether a change worked.Sat, 10 Oct 2026 12:00:00 GMTTrendingKyle Kukshtelhttps://nishtahir.com/build-your-own-decision-model/https://nishtahir.com/build-your-own-decision-model/Nish Tahir shows how to make a language model pick one of a fixed set of answers in a single pass, then tune its confidence to match how often it is right. A GitHub repo has the scripts to try it.Sat, 10 Oct 2026 12:00:00 GMTTrendingNish Tahirhttps://github.com/rociiu/taloryshttps://github.com/rociiu/talorysA free, open-source assistant that lives in your own Cloudflare account and installs with one command. It chats, keeps personal facts you can edit, manages tasks and notes, and sends reminders on a schedule.Sat, 10 Oct 2026 12:00:00 GMTTrendingrociiuhttps://commandline.microsoft.com/microsoft-decision-1-model-foundry/https://commandline.microsoft.com/microsoft-decision-1-model-foundry/Decision models give a short, structured answer that software can act on, instead of text. Microsoft says its new one was the fastest and most accurate in its tests. It is out now in Microsoft Foundry and OpenRouter.Fri, 09 Oct 2026 12:00:00 GMTModelsMicrosofthttps://huggingface.co/papers/2610.13527https://huggingface.co/papers/2610.13527Inside models that read images, the authors found a space that stores each object's position: left-right, up-down and distance. Models build it well but struggle to turn it when the viewpoint changes.Fri, 09 Oct 2026 12:00:00 GMTResearchPaperhttps://huggingface.co/Qwen/Qwen-Image-2.1-Turbohttps://huggingface.co/Qwen/Qwen-Image-2.1-TurboIt makes and edits images in 8 noise-removal steps, using the same 7-billion-parameter design as Qwen-Image-2.1. Anyone can download the weights, which come under the Qwen Research License.Fri, 09 Oct 2026 12:00:00 GMTOpen sourceQwenhttps://arxiv.org/abs/2610.13928https://arxiv.org/abs/2610.13928Researchers hid encrypted circuits in a model's weights that switch on only with a secret trigger. Common ways of drawing out hidden abilities, such as fine-tuning, could not reveal them, even with the weights.Fri, 09 Oct 2026 12:00:00 GMTTrendingPaperhttps://momo5502.com/posts/2026-10-09-game-decompilation/https://momo5502.com/posts/2026-10-09-game-decompilation/Over months, Claude and Codex agents turned a popular first-person shooter back into source code. Maurice Heumann says 99% of its functions were rebuilt, 83% byte for byte, using an estimated 600 to 700 billion tokens.Fri, 09 Oct 2026 12:00:00 GMTTrendingMaurice Heumannhttps://huggingface.co/tencent/Youtu-Parsing-Omnihttps://huggingface.co/tencent/Youtu-Parsing-OmniA 5-billion-parameter model that turns a document page, photo, chart, sound clip or video into one structured data file listing its text, tables, timestamps and a description. The weights are on Hugging Face.Fri, 09 Oct 2026 12:00:00 GMTOpen sourceTencenthttps://github.com/franzenzenhofer/big-arrow-on-the-screenhttps://github.com/franzenzenhofer/big-arrow-on-the-screenA free macOS command-line tool, with a skill for Claude Code and Codex, that lets an agent draw an arrow and a note over every window to show you what to click. Clicks pass through, and the arrow removes itself.Fri, 09 Oct 2026 12:00:00 GMTTrendingfranzenzenhoferhttps://www.cockroachlabs.com/blog/experiment-running-hospital-code/https://www.cockroachlabs.com/blog/experiment-running-hospital-code/For five months, agents in hospital-style roles planned, reviewed and tested changes to its database migration tools. It says over a million lines of code needed only 7 reverts, and IBM Db2 support took under two days.Fri, 09 Oct 2026 12:00:00 GMTTrendingCockroach Labshttps://www.primeintellect.ai/blog/prime-agent-rusthttps://www.primeintellect.ai/blog/prime-agent-rustOver two weeks, Prime Intellect's coding agent ran more than 2,000 agents that rewrote it from TypeScript into Rust. The company says it now runs faster and uses fewer resources than most coding agents.Fri, 09 Oct 2026 12:00:00 GMTTrendingPrime Intellecthttps://www.anthropic.com/news/anthropic-cyber-missionhttps://www.anthropic.com/news/anthropic-cyber-missionAnthropic's new OSS Scanner gives open-source projects regular security checks from its strongest models for free. A second program brings Claude and Anthropic engineers to firms that guard power grids and water systems.Thu, 08 Oct 2026 12:00:00 GMTModelsAnthropichttps://huggingface.co/papers/2610.12572https://huggingface.co/papers/2610.12572A way to fine-tune image and video generators with reward scores, with no per-task tuning. On a 13B video model the authors say it matched two rival methods using 12.02 and 11.72 times fewer GPU hours.Thu, 08 Oct 2026 12:00:00 GMTResearchPaperhttps://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026https://cloud.google.com/blog/products/ai-machine-learning/welcome-to-gemini-at-work-2026Google Cloud says its new Gemini agent is one assistant that answers questions, does office tasks, makes images and media, and writes and runs code, including inside Gmail, Docs, Sheets and other Workspace apps.Thu, 08 Oct 2026 12:00:00 GMTModelsGoogle Cloudhttps://claude.com/resources/articles/dashboards-and-motionhttps://claude.com/resources/articles/dashboards-and-motionAnthropic's Claude Dashboards turns company data into charts that update as the data changes, and Claude Motion makes short animations. Both are in beta; Claude Docs, Slides and Design now work on every plan.Thu, 08 Oct 2026 12:00:00 GMTModelsAnthropichttps://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-Arabichttps://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-ArabicVoxtral Mini 4B Realtime Arabic writes down Arabic speech as it is spoken, in dialects and Modern Standard Arabic, even when speakers switch languages. Made with Morocco's digital ministry; Apache 2.0 licence.Thu, 08 Oct 2026 12:00:00 GMTOpen sourceMistral AIhttps://huggingface.co/blog/lightonai/lightonocr-3https://huggingface.co/blog/lightonai/lightonocr-3OCR turns pictures of pages into text. LightOn's three new models, from 0.8 to 4 billion parameters, can also box and label each part of a page and turn charts into tables. All use the Apache 2.0 licence.Thu, 08 Oct 2026 12:00:00 GMTOpen sourceLightOnhttps://huggingface.co/papers/2610.12374https://huggingface.co/papers/2610.12374Simulators keep each world's rules and state while a video model draws what the agent sees, in real time. The authors say agents learned in 4 rounds, against millions for usual trial-and-reward training.Thu, 08 Oct 2026 12:00:00 GMTResearchPaperhttps://huggingface.co/papers/2610.11765https://huggingface.co/papers/2610.11765Researchers asked six language models for bare answers in six different ways. The models still wrote out visible reasoning, more so on open-ended questions, a habit the paper calls Thinking Inertia.Thu, 08 Oct 2026 12:00:00 GMTResearchPaperhttps://machinelearning.apple.com/research/normalizing-trajectory-modelshttps://machinelearning.apple.com/research/normalizing-trajectory-modelsImage generators usually remove noise over many small steps. Apple researchers say their NTM method matches or beats strong image models in four steps and can still score exactly how likely each result is.Thu, 08 Oct 2026 12:00:00 GMTResearchApple Machine Learning Researchhttps://jessewaites.com/blog/post/i-pointed-ai-at-400-years-of-archives/https://jessewaites.com/blog/post/i-pointed-ai-at-400-years-of-archives/An engineer used Claude Code and other AI tools to search millions of old Dutch East India Company pages and newspapers. The author says this found a forgotten meteorite report, lost rhinos and unrecorded eruptions.Thu, 08 Oct 2026 12:00:00 GMTTrendingJesse Waiteshttps://huggingface.co/papers/2610.08215https://huggingface.co/papers/2610.08215Text games with unfamiliar rules make agents learn by playing. Keeping full records of past moves worked better than boiling them down into rules, and top human players still reached higher peak scores.Thu, 08 Oct 2026 12:00:00 GMTResearchPaperhttps://github.com/pingdotgg/ts-rusthttps://github.com/pingdotgg/ts-rustIts author says Claude Opus 5.5 rewrote TypeScript's compiler, type checker and editor server in Rust for about $24,000 of API use, after OpenAI models spent over $400,000 and stalled near 84% compatibility.Thu, 08 Oct 2026 12:00:00 GMTTrendingpingdotgghttps://www.anthropic.com/claude-haiku-5-5https://www.anthropic.com/claude-haiku-5-5Anthropic says its new small model is its fastest so far and costs around 75% less to run than Haiku 4.5. It is also the first Haiku model with an adjustable effort setting.Wed, 07 Oct 2026 12:00:00 GMTModelsAnthropichttps://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3Bhttps://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3BMiniCPM-V 4.7 understands images and video. OpenBMB says it writes over 13 tokens (word pieces) a second on an iPhone 17 Pro Max. The weights are free under the Apache 2.0 licence.Wed, 07 Oct 2026 12:00:00 GMTOpen sourceOpenBMBhttps://blogs.windows.com/windowsdeveloper/2026/10/07/microsoft-execution-containers-policy-driven-containment-for-ai-agents/https://blogs.windows.com/windowsdeveloper/2026/10/07/microsoft-execution-containers-policy-driven-containment-for-ai-agents/Microsoft Execution Containers runs an AI agent, or just the code and tools it uses, inside a sandbox on Windows, macOS or Linux. Developers list the files and network addresses it may reach; MXC blocks the rest.Wed, 07 Oct 2026 12:00:00 GMTOpen sourceMicrosofthttps://huggingface.co/blog/LiquidAI/open-d1https://huggingface.co/blog/LiquidAI/open-d1Decision models pick an answer in one step instead of writing text. Liquid AI says d1-3B reads text and images and answers in 16 milliseconds on an NVIDIA Jetson AGX Thor board.Wed, 07 Oct 2026 12:00:00 GMTOpen sourceLiquid AIhttps://www.microsoft.com/en-us/research/blog/agent-lightning-v1-0-a-3500-line-lightweight-agentic-rl-framework-for-training-agents-with-real-harnesses/https://www.microsoft.com/en-us/research/blog/agent-lightning-v1-0-a-3500-line-lightweight-agentic-rl-framework-for-training-agents-with-real-harnesses/A framework of about 3,500 lines of code that improves AI agents with reinforcement learning (learning from rewards) while they keep running in their usual tools, with no changes to that code.Wed, 07 Oct 2026 12:00:00 GMTOpen sourceMicrosofthttps://research.google/blog/does-better-work-always-mean-better-workers/https://research.google/blog/does-better-work-always-mean-better-workers/In a three-month trial with 133 patent lawyers, Google researchers found AI lifted the quality of everyone's work, but only senior lawyers came out with stronger judgment of their own.Wed, 07 Oct 2026 12:00:00 GMTResearchGoogle Researchhttps://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/Anyone can now use it to check whether an image, video or audio file carries SynthID, Google's hidden watermark, meaning AI from Google or partners such as OpenAI and NVIDIA made it. It works worldwide, in English.Wed, 07 Oct 2026 12:00:00 GMTModelsGoogle DeepMindhttps://blogs.nvidia.com/blog/local-ai-rtx-spark-microsoft-windows-event/https://blogs.nvidia.com/blog/local-ai-rtx-spark-microsoft-windows-event/A chip with up to 128 GB of memory shared by its graphics and main processors, so laptops and small desktops can run AI models locally. NVIDIA says preorders open today.Wed, 07 Oct 2026 12:00:00 GMTModelsNVIDIAhttps://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026NVIDIA says tuned versions of Nemotron 3 scored at gold-medal level in this year's maths and coding olympiads (IMO and IOI; the coding run was unofficial), and has shared the models and training data on Hugging Face.Wed, 07 Oct 2026 12:00:00 GMTOpen sourceNVIDIAhttps://huggingface.co/papers/2610.12242https://huggingface.co/papers/2610.12242Routing splits work between language models; doing it for every token (word piece) can be cheaper and better but is hard to serve. TokenRouter handles 2 to 64 times more output per second than existing systems.Wed, 07 Oct 2026 12:00:00 GMTResearchPaperhttps://github.com/openai/math/blob/main/history.mdhttps://github.com/openai/math/blob/main/history.mdA sign error in one proof also broke two papers that relied on it, so OpenAI withdrew all three. It repaired 14 others and says 300 of the collection's 719 main results are now formalized (checked by proof software).Wed, 07 Oct 2026 12:00:00 GMTTrendingOpenAIhttps://claude.com/resources/articles/claude-now-works-in-google-docs-sheets-and-slideshttps://claude.com/resources/articles/claude-now-works-in-google-docs-sheets-and-slidesClaude opens in a sidebar next to your file to fix text, write formulas, build charts and pivot tables, or add slides that match your deck. It is in public beta on all paid Claude plans.Tue, 06 Oct 2026 12:00:00 GMTTrendingAnthropichttps://mistral.ai/news/mistral-large-4/https://mistral.ai/news/mistral-large-4/Mistral says its largest model yet has 1 trillion parameters, 49 billion of them active at a time, and understands images as well as text. A preview API is open now; Mistral plans to release the weights by month's end.Tue, 06 Oct 2026 12:00:00 GMTModelsMistral AIhttps://huggingface.co/papers/2610.08448https://huggingface.co/papers/2610.08448The student learns from a teacher model's feedback on its own answers. The authors find that comparing only the student's top 16 shared tokens (word pieces) works about as well as comparing them all.Tue, 06 Oct 2026 12:00:00 GMTResearchPaperhttps://huggingface.co/papers/2610.08077https://huggingface.co/papers/2610.08077When every practice run of an AI agent gets the same score, reward-based training has nothing to learn from. SRD trains the agent to foresee what finished runs reveal; in one test, success rose from 0% to 60.6%.Tue, 06 Oct 2026 12:00:00 GMTResearchPaperhttps://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/A 740-million-parameter model built to run on the device itself. It turns text, images, audio and video into embeddings (lists of numbers that capture meaning) for search and retrieval. Licence: Apache 2.0.Tue, 06 Oct 2026 12:00:00 GMTOpen sourceGooglehttps://machinelearning.apple.com/research/rised-multi-environment-selectionhttps://machinelearning.apple.com/research/rised-multi-environment-selectionAn AI judge labels each attempt with shared tags for good and bad behaviour. RISED uses the tags to choose training data, coach the model and steer it away from repeated mistakes.Tue, 06 Oct 2026 12:00:00 GMTResearchApplehttps://research.google/blog/earth-ais-planetary-geospatial-foundation-models-for-global-public-health/https://research.google/blog/earth-ais-planetary-geospatial-foundation-models-for-global-public-health/Google Research shares five case studies in which partners used PDFM, a model that sums up each place as a list of numbers, to study health questions such as cardiovascular disease deaths in U.S. counties.Tue, 06 Oct 2026 12:00:00 GMTResearchGoogle Researchhttps://research.google/blog/open-and-emergent-problems-in-agentic-privacy-and-security-a-contextual-angle/https://research.google/blog/open-and-emergent-problems-in-agentic-privacy-and-security-a-contextual-angle/Google Research shares a report by more than 50 experts from universities and companies on the privacy and security problems of AI agents. It argues agents must learn what is appropriate in each situation.Mon, 05 Oct 2026 12:00:00 GMTResearchGoogle Researchhttps://huggingface.co/papers/2609.39601https://huggingface.co/papers/2609.39601This 4-billion-parameter model marks objects in images with points and boxes. Across 34 tests, its average beats the larger GPT-6 Astra, and it helps robot arms and self-driving systems as their vision part.Wed, 30 Sep 2026 12:00:00 GMTResearchPaperhttps://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/Google's new frontier model, built for long, complex work. For now it is rolling out only to trusted cyber defenders; Google says wider access for developers and businesses will follow.Wed, 30 Sep 2026 12:00:00 GMTModelsGooglehttps://blog.janestreet.com/can-you-use-autoregressive-diffusion-to-generate-market-data/https://blog.janestreet.com/can-you-use-autoregressive-diffusion-to-generate-market-data/A Jane Street intern built a diffusion model that invents order-book events, the orders traders place or pull. Flow matching trained far better than standard DDPM, but the model is not yet realistic enough.Wed, 30 Sep 2026 12:00:00 GMTTrendingJane Streethttps://openai.com/index/introducing-gpt-6-1-solhttps://openai.com/index/introducing-gpt-6-1-solA model for coding, computer use and professional work that OpenAI says comes close to GPT-6 Astra at a fifth of Astra's API price.Tue, 29 Sep 2026 12:00:00 GMTModelsOpenAIhttps://github.com/openai/codex/issues/48195https://github.com/openai/codex/issues/48195Since version 0.157, opening OpenAI's Codex command-line tool starts a helper that keeps running after you close it and checks hourly for updates. The issue asks OpenAI to make it opt-in.Fri, 25 Sep 2026 12:00:00 GMTTrendingopenai