Open Source Radar — September 9: video, plugins, and a 2B model

Today's board splits four ways: a renderer that turns agent-written HTML into real video, a skill that cuts the preamble out of agent answers, OpenAI quietly moving Codex's extension surface into one marketplace, and an open-weights 2.5B model that says it beats everything in its weight class.

HyperFrames (TypeScript, ~48,200 stars, Apache-2.0) — HeyGen's open-source renderer takes HTML, CSS, media and seekable animations and outputs a deterministic MP4. That's a much better fit for an agent than a video editor: everything is text the model can write, lint, and diff, and the render is reproducible rather than a screen recording. It ships 20 published skills covering the production loop — plan the piece, write valid HTML, wire the animation, add media, lint, preview, render — with a router skill that picks the workflow for a product-launch clip, a pull-request changelog video, or captions burned into a talking-head take. The daily board added roughly 2,600 stars and releases are landing several times a day (0.8.33 this morning, 0.8.32 and 0.8.31 yesterday), so expect churn. Reach for it when you want your agent to hand back a finished video instead of a script.

i-have-adhd (Python, ~32,400 stars, MIT) — A skill with exactly one job: stop your coding agent from burying the answer. It's ten rules — lead with the next action, number multi-step tasks, end with one concrete next step, cap lists at five, no preamble, no recap, no "Hope this helps!" — and the README runs the same answer before and after. The before is the familiar paragraph that explains your auth system, offers context you didn't ask for, and invites you to dig deeper; the after is three numbered steps, a file path, and one line telling you what to paste back if a test fails. It made the Hacker News front page on September 8 and ships READMEs in Chinese, Portuguese, Japanese, Vietnamese, Korean and Thai, which tells you who's installing it. Install it if your agent's prose is the reason you've stopped reading its answers.

openai/plugins (JavaScript, ~6,000 stars) — The repo OpenAI now points to for Codex extensions, after it deprecated the older skills catalog earlier this year. Each entry is a folder holding a manifest plus optional skills, MCP servers, commands and hooks, and there are 62 of them right now: Figma, Notion, Linear, Stripe, Supabase, Sentry, Cloudflare, Slack, plus build-an-app packs for iOS, macOS and web. Qodo was added on September 8 and CrowdStrike's Foundry and Fusion landed the week before, which is the more interesting tell — security and code-review vendors are treating agent extensions as a distribution channel. Small star count, but this is where Codex's third-party surface actually lives, so it's the repo to watch rather than the deprecated one.

MiniCPM5-2B (Python, ~10,700 stars) — OpenBMB released this 2.5-billion-parameter dense model on September 7, aimed squarely at on-device and resource-constrained deployment: 2.52B parameters across 42 layers in a plain Llama-style architecture, with a 131,072-token context window, and BF16, GGUF, MLX and GPTQ builds already published. The team claims state of the art for its size class — a 53.9 average across the benchmark set it publishes, ahead of Qwen3.5-2B, Gemma-4-E2B and LFM2.5-2.6B, and above the Qwen3.5-4B reference point it lists — with the margin widest in code, math, long-context understanding, tool use and agentic tasks. Early community testing puts it around 15 on Artificial Analysis's index, which would make it the strongest open-weights model at 4B or below; treat that number as a forum reading, not a model card. What makes it a radar pick rather than a release note is the packaging: the training datasets ship alongside it, and so do deployment and fine-tuning agent skills.

Worth watching this week: two of today's four are about what an agent hands back — a rendered file instead of a plan, an answer instead of an essay — and the other two are about where it gets its reach and its brain.

Which of these would you actually install today — a video renderer, a bluntness skill, a plugin marketplace, or a 2B model that runs on your laptop? Tell us in the comments.

Sources: HyperFrames (GitHub) · i-have-adhd (GitHub) · openai/plugins (GitHub) · openai/skills (deprecated catalog) (GitHub) · MiniCPM (GitHub)

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

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Read more

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

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute