Open Source Radar — September 11: skills go open, models go home

The open-source story today is about ownership: who holds the model, who holds the agent's instructions, and who keeps the knowledge after the chat ends. Five picks from the daily board.

Colibrì (C, 27,632 stars) — The most ambitious project on the board: an inference engine, in pure C with zero dependencies, that runs frontier mixture-of-experts models up to 2.8 trillion parameters on hardware you already own. The trick is treating your SSD, RAM, and VRAM as one memory hierarchy — the always-needed parts of the model sit in RAM, and the thousands of routed experts stream from disk on demand, like a JIT compiler that compiles hot paths just in time. The author's demo shows GLM-5.2 (744B parameters) answering at 4 tokens per second from a machine with about 10 GB of resident memory, and the project honestly publishes which optimizations are measured and which are still hypotheses. It's research-grade rather than plug-and-play, and disk speed matters enormously — but if you want to hold a frontier-class open-weights model instead of renting it by the token, this is the most credible attempt we've seen.

OpenMAIC (TypeScript, 35,680 stars) — Tsinghua's open multi-agent interactive classroom: one prompt in, a complete multi-agent course out. Director agents orchestrate the curriculum, then agents generate slides, quizzes, simulations, 3D visualizations, and voice narration, and the classroom plays back as an interactive lesson rather than a wall of text. Version 1.0.0 landed August 27 with an agent workbench — you can now chat with an agent that plans, builds, and revises the whole course, uploading your own documents, audio, and video as source material. It's provider-neutral (bring your own models), MIT-licensed, deployable with one click, and backed by a published paper. If you've been waiting for the "AI tutor" to become something a teacher or a self-learner can actually run, this is the furthest along.

vercel-labs/skills (TypeScript, 31,358 stars) — Skills became the standard way to teach coding agents new tricks; this is the tool that made them portable. One CLI installs any skill from any git repo into any of 75+ supported agents — Claude Code, Codex, Cursor, OpenCode — by symlinking a canonical copy into each agent's skills directory, so one install serves every tool on your machine. There's a discovery layer at skills.sh, a mode that runs a skill without installing it, and project-scoped installs that commit with your repo so your whole team shares the same setup. MIT-licensed, and trending because the skills ecosystem itself is exploding. If you've been copying skill files between agents by hand, this is the fix.

LLM Wiki (TypeScript, 18,404 stars) — Andrej Karpathy described a pattern for a personal knowledge base that an LLM incrementally builds and maintains instead of answering from scratch every time; nash_su built it into a full cross-platform desktop app. It ingests your PDFs, Office files, e-books, web clippings, even images (with vision-model captions), and compiles them into an interlinked wiki with source-traceable pages, a knowledge graph, and community detection that surfaces clusters and gaps. The key design bet is that knowledge is compiled once and kept current, not re-derived per query — and the wiki directory stays Obsidian-compatible, so your data never locks in. A local API and MCP server let your coding agent query the wiki read-only. GPLv3. The closest thing to RAG that actually gets smarter the longer you use it.

OpenClaw integrations are quietly eating the ecosystem — Two of today's five picks ship first-class hooks for OpenClaw, and that's the pattern worth watching: OpenMAIC lists OpenClaw integration as a headline feature, LLM Wiki ships its agent skill with the same one-command installer Vercel's tool standardized. Skills stopped being a Claude Code feature and became the packaging everyone targets. The irony is that the agents themselves are commoditizing while the skills on top are where projects now compete. Worth watching this week: whether the agent-skills specification consolidates fast enough that one skill really does run everywhere.

Building with any of these? Tell us in the comments.

Sources: Colibrì (GitHub) · OpenMAIC (GitHub) · skills (GitHub) · LLM Wiki (GitHub) · Karpathy's llm-wiki pattern

AI 101 — How do AI video generators work?

An AI video generator is a model that starts with a grid of random noise and repeatedly refines it into a moving picture that matches your prompt — the same trick image generators use, stretched across time. Nothing is filmed, and nothing is stitched together from clips the model has seen. The pixels in the finished video were invented, frame by frame, all at once. Why it matters right now Kling 4.0, Kuaishou's flagship video model, went fully live on October 15 with up to 30 seconds of vid

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Read more

AI 101 — How do AI video generators work?

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone