China just dropped the world's most advanced open-source AI.

China just dropped the world's most advanced open-source AI that beat Anthropic and OpenAI for free. And a new tool lets you run a 2.8 trillion parameter model on a 4 gigabytes graphics card. Here's everything that happened in AI this week in under 60 seconds. Number 6, CopilotKit launches OpenBot, an open source Grockbot alternative you can self host. Every bot gets its own computer with its own files and logins. You watch it work live and can take over anytime. Number 5, Alibaba open sources OpenSandbox, the isolated environment your agents actually need. They can run code, browse the web, control a full desktop, or run Claude Code, Cursor, Codex, and Gemini CLI inside it. Number 4, Ripley launches free AI mode, powered by GPT-5.6 Luna. Chat, plan, and iterate all day without burning credits. It only escalates to the expensive model when the task needs it, and it warns you first. Number 3, a new open source tool stops your coding agent from relearning your entire code base every session. It's called Graft, and it maps your code base once into linked markdown files that live in your repo. Up to 4 times fewer tokens, two commands to install, and free. Number 2, Z.ai drops GLM 5.3, the highest scoring open model ever on Terminal Bench. And it just beat Anthropic and OpenAI's best on cybersecurity. And at number 1, AirLLM lets you run a 2.8 trillion parameter model on under 4 gigs of VRAM. It loads one layer at a time, so the VRAM you need depends on layer size, not model size. Completely free and open source. Want to check them out? Comment AI and I'll send you the link to try them all.