Andrew Ng open-sources OpenWorker, a local agent that ships work
Runs locally on your Mac, model-agnostic, checks in before it acts. Plus BFL's multimodal FLUX 3, Codex multi-folder projects, and 4-bit diffusion in Diffusers.
Copy markdown
OpenWorker hands back the deliverable, not a chat log
Andrew Ng released OpenWorker, an open-source agent that runs locally on your Mac (Windows soon) and produces finished work — a drafted doc, a sent Slack message, an untangled calendar — across your files and everyday tools, checking in before anything consequential. It's model-agnostic: point it at Claude, GPT, Gemini, open weights, or a local Ollama model.
FLUX 3 adds 20-second video and audio to one model
Black Forest Labs unveiled FLUX 3, a multimodal model that generates images plus up to 20-second video with synced audio — and action prediction — from a single system, its first step beyond still images. It's a limited release to start, so access is gated, but it stakes out a unified image/video/audio frontier.
Codex projects now span multiple folders
OpenAI's Codex app (build 26.715) lets one local project pull in code, docs, and reference files from several folders, with a designated primary folder as the Git root that drives new chats and AGENTS.md/skills/config.toml discovery. If you've juggled a repo plus scattered reference dirs, they're one project now.
4-bit diffusion lands natively in Diffusers
Hugging Face folded Nunchaku's SVDQuant 4-bit inference into Diffusers, so a from_pretrained() call runs FLUX-class image models at about half the VRAM (31.1GB to 16.0GB peak) and ~1.35x faster — up to 1.8x with torch.compile — with no separate engine or ComfyUI. INT4 covers Turing/Ampere/Ada; NVFP4 needs Blackwell.