Open Source Radar — November 9: Meta's Astryx and Tencent's 770B Hy4
Today's board is a design system, a GPU control plane and Anthropic's reference apps — while the model everyone is actually waiting on is Tencent's Hy4, still arriving in pieces.
Hunyuan Hy4 (MoE, 770B total parameters and 49B active, 1M-token context) is Tencent's bid for the open-weights frontier, and the part that matters today is timing. Hy4 shipped on November 4 with a promise attached: the full weights and a system card land in the coming days, with vLLM and SGLang support lined up alongside them, while the August preview is already Apache 2.0 on Hugging Face — so you can run the architecture now and swap the files later. What is new in the shipping version is the agent surface: first native vision and first computer use in the line, trained on Blender, CAD and multi-app workflows, which is the difference between an open model that talks about your screen and one that can drive it. QbitAI's hands-on this morning is the useful read — given a screenshot of a music-player design it rebuilt the page, wrote its own headless-browser tests, caught a broken next-track edge case and passed all 59 checks on the rerun.
Astryx (TypeScript, about 18,000 stars, MIT, beta) is the design system Meta has run internally for eight years — 13,000+ apps, the company's largest — rebuilt on React 19 and StyleX and published in June. It ships 150+ accessible components as pre-compiled CSS, so there is no build plugin to adopt; a theme is a set of CSS variable overrides rather than a fork, and your own class names layer on top, which means a Tailwind project can use it without giving up Tailwind. It is trending today on pace rather than novelty: 7,151 commits, release v0.7.11 two days ago, and a fix committed ten minutes before this was written. The agent part is the pitch worth stealing — the CLI and an MCP server expose the same docs an engineer reads, so the assistant builds with the same API instead of guessing at it.
NVKE (Go, 978 stars, Apache-2.0, preview) is NVIDIA's management plane for selling Kubernetes on your own GPUs: you install it on bare metal and it provisions machines through NVIDIA's Infra Controller, then stands up independent workload clusters, each with its own control plane, etcd and node pools, driven by an API or the nvkectl command line. Nodes attest through SPIFFE/SPIRE with their TPM before they receive join credentials, and a companion repository builds the node images from pinned inputs — the difference between a cluster someone assembled once and one you can reissue on demand. It is explicitly a preview and not recommended for production, so the honest entry point is the local quickstart that simulates machines on a Linux box with 6 CPUs and 12 GiB of RAM and no GPU at all.
claude-quickstarts (Python, about 19,700 stars, MIT) is the least glamorous item here and probably the most reused: Anthropic's reference apps for computer use, autonomous coding, a browser toolset, a customer-support agent and a financial analyst, all pinned to the current SDK. The newer folders carry the signal — Managed Agents examples for a Sentry autofix bot and a Slack bridge that keeps no database and no long-lived connection, plus a knowledge wiki rebuilt on a dynamic workflow on November 6. Read it as an opinion about how agents should be deployed rather than as samples: short sessions, webhooks instead of sockets, state in session metadata. If you are building your first agent product, the folder layout alone saves a week of architecture argument.
Worth watching this week: whether Tencent actually posts the full Hy4 weights, and whether Astryx's agent-facing CLI becomes the default reason to pick a design system.
Which of these would you put in front of a paying customer first — an open 770B with computer use, or a design system your agent already knows? Write to us via our contact page.
