Open-Source AI

Today's Stories

1. 1

   Broadcom Unveils VMware AI Factory for Faster Private AI

   Broadcom announced VMware AI Factory, a software-defined foundation for VMware Private AI Cloud, at VMware Explore 2026. It automates AI infrastructure deployment and lifecycle management. This cuts the time from bare metal server deployment to serving the first AI model from weeks to hours. It also pools GPU resources so teams can run multiple models on shared hardware, lowering costs and improving tokenomics.

   Broadcom is partnering with MetalSoft to reduce bare-metal provisioning from weeks to minutes. It also collaborates with AMD to support AMD Instinct GPUs and the open ROCm software ecosystem.

2. 2

   OpenClaw 2.0 launches with major update

   OpenClaw launched version 2.0, its largest update yet, with a version number of 2026.8.1. The update focuses on quality-of-life improvements, better persistence, easier installation, and increased security. OpenClaw is a free, open-source AI agent that acts like a digital secretary, executing tasks on a user's behalf rather than just answering questions like a traditional chatbot. This update removes a prior limitation by enabling shared cloud sessions, allowing team members to collaborate on the same AI agent in real time.

   The team said it now uses this multiplayer collaborative feature to build the AI agent framework code itself. This is the first major version update since OpenAI Group PBC hired Peter Steinberger, the creator of the AI agent framework.

3. 3

   DHH Releases Omarchy 4.0 with Built-in AI Agent

   David Heinemeier Hansson (DHH), creator of Ruby on Rails, has released Omarchy 4.0 (Omarchy Quattro), the latest version of his Linux desktop OS. The release was announced on August 14, 2026. Omarchy now integrates AI agents such as Claude, Codex, Copilot, Gemini, and Grok directly into the OS. The agents can answer questions, install apps, change settings, and even create themes or plugins on request, reducing the need for manual configuration and troubleshooting.

   This version adds support for dual-booting with Windows, and a community tool "Try Omarchy for Windows" lets users run the full OS on Windows without installation, lowering the barrier for Windows users to try it.

4. 4

   Debian allows AI-assisted code contributions

   Debian voted to allow developers to use AI tools in contributions to the Linux distribution, covering development, maintenance, and documentation. The new policy neither endorses nor prohibits generative AI use. The decision rejects proposals that would have banned AI-generated contributions. Some contributors are unhappy, with one saying he's quitting and is 'no longer interested in anything coming from Debian.'

   Debian still holds contributors responsible for AI-assisted work, requiring them to review and test it. Disclosure of AI use is encouraged but not required.

5. 5

   OpenClaw 2.0 simplifies setup, adds cloud sessions

   The OpenClaw Foundation released version 2.0 of its open-source AI platform, the largest release to date with over 16,000 pull requests. It adds shared cloud sessions, a streamlined onboarding process, and a rebuilt browser app. The new setup detects existing resources like ChatGPT or Claude subscriptions, API keys, and local models, cutting many configuration steps. The browser app opens a conversation immediately, and the Session Rail shows progress without interrupting the agent.

   Shared Cloud Sessions let multiple users work on the same tasks, pulling in team members with context. Sessions can run on the local gateway, paired devices, or rented machines via Crabbox with backends like AWS and Hetzner.

6. 6

   Marker v2: 23.7 pages/sec on one B200 GPU

   Datalab released Marker v2, an open-source parsing pipeline that processes up to 23.7 pages per second on a single B200 GPU. It outputs Markdown, JSON, or HTML and runs on Surya 2, a single 650M parameter model handling OCR, layout, reading order, and table recognition across 90+ languages. Traditionally, scaling parsing meant adding workers, each loading its own model copy, leading to idle GPU time and paying for multiple copies while getting throughput of one. Marker v2 instead uses many lightweight CPU workers sharing one Surya inference server, so pages batch together, keeping the GPU busy and throughput scaling with the server, not VRAM.

   The trade-off modes matter: balanced (highest quality, best on GPU), fast (cheaper per page but math accuracy drops sharply because equations read from text layer), and disable_ocr (CPU-only for clean digital files). On Allen AI's olmOCR-bench (1,403 PDFs), balanced scores 76.0% versus MinerU's 72.7% while running over 5x more pages per second.