Google Open Source

Google Open Source

@opensource.google

Announcing new open source releases, exploring projects, sharing how we approach FOSS, and supporting communities around the world.

3.5K followers83 following475 posts

Pinned

Open source is free as in puppies, not free as in beer. 🐶📦 While research often looks at upstream maintenance, @sophiaiv.bsky.social studied the effort to USE open source software at scale across Google. What makes updates take days vs. hours? Dive into the data: goo.gle/cost-of-oss

Google Open Source reposted

467 providers serve a conformant A2A agent card. Nearly half came from a docs platform default with a single skill: read our docs. Most of the rest are agent-native companies selling company data, trust and on-chain services by the call. Conformant is not the same as useful.

apievangelist.comWhat 467 Conformant A2A Agent Cards Are Actually OfferingAs of today’s build, the A2A section of APIs.io has 594 providers serving an agent card. 467 of them are conformant, meaning the card passes every structural...

How do you test and serve Agent2Agent (A2A) v1.0 agents directly from the terminal without writing custom SDK boilerplate? @meteatamel.bsky.social walks through the official A2A CLI (`a2a card get`, `a2a send --stream`, `a2a server --echo`/`--exec`, and `a2a skill`): https://goo.gle/a2a-cli-guide

goo.gleTesting and Serving A2A Agents from the Terminal with A2A CLIMete Atamel walks through testing and serving Agent2Agent (A2A) v1.0 agents from the terminal using the official a2a CLI (a2a card get, a2a send --stream, a2a server --echo/--exec, and a2a skill).

🌐Join us at OSSummit Japan + AutoLinux + EmbeddedLinux Conference Asia, December 7-9 in Tokyo! We’re proud to support the OpenSource community as a Diamond sponsor of this can’t-miss event. Explore the schedule ➡️ goo.gle/4saJcId 🎟️ Register today!

Heading to KubeCon + CloudNativeCon NA? Visit the Community Hub! 🤝✨ Connect with peers, join interactive workshops, and celebrate inclusion with @cncf.io and @linuxfoundation.org. What sessions are you looking forward to? Learn more: goo.gle/kubecon-na-c...

goo.gleCommunity Hub | KubeCon + CloudNativeCon North AmericaThe Community Hub is a welcoming space for attendees to connect, learn, and celebrate equity, inclusion, and accessibility. The Hub will host community gatherings in collaboration with Merge Forward…

This Week in Open Source #20 is live! 🌐✨ 🏛️ OSI at UNGA81 & “AI-DDoS” maintainer research (-18.18% newcomer merge rate) 🔍 RedMonk’s 24K commit audit & OSS-Fuzz-inspired security scanning 🎙️ Gemma 4 4B DiarizationLM, skill-reach & offline ESP32-S3 voice AI: goo.gle/twios-20

Google Open Source reposted

Instead of going token by token, DiffusionGemma refines a 256-token block at once on the same base model as Gemma 4. In my 10-task test, it did better on copy editing and upfront counts, missed on valid JSON and reasoning, and ran slower on my Mac but much faster on a GPU.

medium.comWhen to use a diffusion LLM: I compared DiffusionGemma and Gemma 4DiffusionGemma 26B and Gemma 4 26B share a base model but decode differently. Here's where block diffusion wins, where it breaks, and how speed scales.

Huge milestone for the open social web—the first dedicated AT Protocol devroom is coming to FOSDEM 2027 in Brussels (Jan 30–31)! Congrats to @viivi.rocks, @edouard.paris, and the @atproto.eu community on bringing the room together. Watch for the CfP by Oct 27!

Viivi@viivi.rocks

AT Protocol devroom is coming to FOSDEM 2027 🎉 @edouard.paris and I will open the CfP by October 27, mark your calendars ✍️

Keeping data open is just as important as keeping source open. Source code, it could be argued, is data anyway.

Processor press@processorpress.bsky.social

Google and the United Nations announced the UN System Data Commons, an open‑source platform that brings together global statistics from UN agencies into a single, searchable hub. The launch was highlighted on… #Google #Launch #Open #Data #Commons #Simplify #InternationalNews #Explained #AIWire

Google Open Source reposted

Google Cloud, Red Hat and IBM release llm-d 1.0, a stable open-source inference stack. It routes requests to servers that already hold the matching cached prompt and splits prompt processing from token generation onto separate machines.

Kicking off Day 2 @ 9:15 AM with a Keynote from Bill Jia, VP of Eng, Core ML/AI, Google Cloud! 🚀 Explore PyTorch hardware heterogeneity, the TorchTPU OSS release, and how agentic workflows simplify scaling AI workloads across any silicon. goo.gle/4Aa3Mw9

What was your favorite part of All Things Open this year? 🗳️ 🔹 Keynotes & Tech Talks 🔹 Networking in the Community Lounge 🔹 Expo Hall & Swag 🔹 Connecting with Open Source Devs Let us know below! 👇

Google Open Source reposted

Google Open Source reposted

Google Research released Harness², an open-source framework that tunes an AI agent's harness — prompts, tools, memory, guardrails — instead of the model. First-party tests show gains up to 11.5 points. The catch: more runs, more API cost and no sandbox by default. techstrong.ai/agentic-ai/g...

techstrong.aiGoogle Research Releases Harness² to Improve AI Agents Without Retraining ModelsBigger models aren't the only route to smarter AI agents. 🧠Google's open-source Harness² lets agents improve their own prompts, memory, tools and workflows—without retraining.Early gains reach 11.8 po...

Google Open Source reposted

unite.aiGoogle Open-Sources TorchTPU, Putting Native PyTorch on Its TPUsGoogle announced on October 20, 2026, at PyTorch Conference 2026 that TorchTPU, its native PyTorch stack for Tensor Processing Units (TPUs), is now available as open source on GitHub, the company said on the Google Cloud blog. The software was co-developed with the PyTorch team at Meta and is designed to run existing PyTorch workloads on TPUs with minimal code changes. With the release, TPUs join GPUs and CPUs as a native option for PyTorch, Google said, spanning training, serving, and deployment through tools including vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, Hugging Face, and LiteRT. Developers can debug models in eager mode and then move to production-level performance through torch.compile to optimize TPU utilization and reduce memory overhead. Jana van Greunen, Director of PyTorch Engineering at Meta, said in the announcement that TorchTPU allows Meta to co-develop hardware flexibly with its core machine learning systems team while supporting major LLM, Ads, and Recommendation systems workloads. “It enables us to deploy our PyTorch workloads to Google TPUs, provide direct feedback to the engineering teams, and build the ecosystem flexibility everyone needs,” she said. Hugging and Serving Integrations Google partnered with Hugging Face so that open-weight models on the Hugging Face Hub...

Join Sandeep Pokkunuri & Chris Jones on Oct 20 at 11:45 AM for "Cloud TPU Nexus: Autonomous Multi-Agent Swarms for PyTorch"! Learn to streamline PyTorch-to-TPU migrations and achieve >70% of hand-tuned performance in <24 hours. goo.gle/4gTLUhw

Join Rob Mulla at 10:40 AM for our sponsored demo: Unifying & Accelerating Open-Source LLM Serving on Google Cloud TPUs with TorchTPU! See how it provides a unified backend for vLLM & SGLang, deploying SOTA models with minimal code changes. goo.gle/4gTLUhw

Google Open Source reposted

Generative AI is reshaping open source development, and the trade-offs are starting to show. LF Research surveyed 234 OSS contributors and maintainers: 1 in 4 incoming PRs are now "AI slop," and only 1 in 3 projects have a formal AI policy. Read the full report to learn more.

linuxfoundation.orgNavigating Open Source Software Development in the Generative AI EraNavigating Open Source Software Development in the Generative AI Era

PyTorch Conference 2026 is live! We’re kicking off 2 days of open source innovation! Stop by booth D5 to chat with our team & grab some swag! 🌐✨ Which sessions are on your radar today?

Google Open Source reposted

Web Tiles in the news!

Google Open Source@opensource.google

This Week in Open Source #19 is live! 🌐✨ 🐚 A2A + MCP from Toronto to your terminal (A2A CLI & a2term) 🧲 Local multimodal search w/ EmbeddingGemma 2 (567MB RAM) 🧩 Live AT Protocol 3D Web Tiles 🐧 35 years of Linux & 2026 State of OSPOs w/ @linuxfoundation.org: goo.gle/twios-19

Stop by the Google Community Lounge at All Things Open! ☕ It’s the place to recharge, network, and chat with members of the @FlutterDev team. Drop by to say hi and talk all things open source!

"Open source projects work best when they are close to the users." @emckean.bsky.social spoke with @thenewstack.io on Docsy joining @linuxfoundation.org, llms.txt, and agent-readability doc scores. How are your docs adapting? thenewstack.io/docsy-linux-...

thenewstack.ioWhat's up, Docsy? Google’s docs project joins the Linux Foundation as AI agents become readersDocsy already generates Markdown and llms.txt for AI tools. Next, the open source project wants to score how easily agents can read your docs.

This Week in Open Source #19 is live! 🌐✨ 🐚 A2A + MCP from Toronto to your terminal (A2A CLI & a2term) 🧲 Local multimodal search w/ EmbeddingGemma 2 (567MB RAM) 🧩 Live AT Protocol 3D Web Tiles 🐧 35 years of Linux & 2026 State of OSPOs w/ @linuxfoundation.org: goo.gle/twios-19

Joining AGNTCon + MCPCon NA this Oct 22-23 in San Jose! The agenda covers building & scaling agentic AI systems in production—from orchestration to evaluation, resilience, and debugging. See you there! Schedule ➡️ goo.gle/4AacRov

Big fans of citizen science here at Google Open Source. Powering it with Gemma 4 is even more exciting.

Baltar@balthazar42.bsky.social

youtube.com/watch?v=dxq-... And there is a thing called citizen science. I've discussed it with Gemini, and you can use the Google AI Gallery Edge, the offline AI provided by Google that uses Gemma 4 and other models. You

Google Open Source reposted

Google’s MCP Toolbox Java SDK separates agent reasoning from controlled data access: the model selects tools while predefined queries and bound parameters constrain execution. Let AI decide intent. Let deterministic systems enforce authority. cloud.google.com/blog/topics/... #MCP #Java #AgenticAI

cloud.google.comAnnouncing MCP Toolbox Java SDK v1.0: Agentic data access for the enterprise | Google Cloud BlogDiscover MCP Toolbox Java SDK v1.0: secure, type-safe agent orchestration for enterprise Java, with decoupled auth and AlloyDB & Spring Boot support.

"Open by itself does very little, but enables a lot." At Open @ UNGA81 with @opensource.org & @creativecommons.bsky.social, our own amanda casari explored what it takes to make good on openness—from investing in vendor-neutral standards to rethinking trust & maintainer tooling in the AI era.

Open Source Initiative@opensource.org

At UNGA81, OSI joined conversations on Open Source AI, digital sovereignty, and digital public infrastructure. Open by itself does very little but enables a lot.... https://opensource.org/blog/open-enables-a-lot-the-osi-at-unga81

🗓️ The countdown to All Things Open is on! We’re proud to sponsor this year’s event. Join thousands of open source developers, creators, and leaders for world-class talks, networking, and innovation. Grab your ticket & see us there: 2026.allthingsopen.org

The past few days have been a blur. Open Source Summit + Embedded Linux Conference Europe was packed with news and information. What open source tidbits have you the most excited?

Check out our OSPO's own @fabricatedtechnobabble.com's write up of his talk and experience at MCP Dev Summit in Toronto earlier this week.

Daryl Ducharme@fabricatedtechnobabble.com

What a week at #MCPDevSummit Toronto. I presented my talk "Tag-Team Transmission: Navigating A2A and MCP for Optimum Orchestration". Full recap, photos & slides on why treating agents like tools breaks in production: fabricatedtechnobabble.com/a/3mxc6rqulr...