Open Source Radar — November 22: Train a 64M LLM for $0.40

Share
Open Source Radar — November 22: Train a 64M LLM for $0.40

Three repos trending today point the same way: models small enough to own outright, and agents finally getting a way to operate software that was never built for them.

MiniMind (Python, 68.3k stars) is a complete from-scratch training pipeline for a 64-million-parameter language model, and the number that matters is the price. One epoch of supervised fine-tuning on a single RTX 3090 takes about 2.3 hours of rented GPU time, roughly three yuan — around 40 US cents, per the project's own benchmark table. The model itself is far too small to be useful for real work; the code is the product. Pretraining, SFT, LoRA, DPO, PPO/GRPO, tool use, agentic RL and distillation are all written in plain PyTorch, with none of the transformers/trl/peft abstractions that hide the mechanics. If you have wanted to see how an LLM is actually built, this is the cheapest route from zero, and the MoE variant at 198M total (64M active) shows the routing machinery alongside it.


CLI-Anything (Python, 62.3k stars) answers a dull but expensive problem: agents are fluent in command lines and helpless in graphical apps. The HKU Data Intelligence Lab's generator walks a piece of software — GIMP, Blender, FreeCAD, QGIS, LibreOffice, Krita, Zotero, Obsidian, Godot — and emits a Python CLI harness with structured JSON output plus a skill file, so an agent that cannot click a menu can still resize an image or author a 3D scene. A companion registry installs ready-made harnesses, 258 commands for FreeCAD alone, and the project reports 2,464 passing tests across 18 demo apps, according to its test results. For anyone wiring an agent into a real workflow, the harness layer is where most of the hidden work lives.


AutoClip (13.1k stars, MIT) is the open-source answer to Opus Clip. Paste a link to a podcast, interview or lecture, pick a destination, and it cuts the long video into vertical shorts with burned-in subtitles, a cover and posting copy for TikTok, Reels, Shorts, Douyin, Xiaohongshu or Bilibili. What makes it more than a wrapper: cutting, reframing and rendering happen on your own machine, transcription can run locally through Whisper or SenseVoice, and the analysis step can point at Ollama or LM Studio instead of a paid API, per the project's documentation. It also ships a CLI and an MCP server, so an agent can trigger the same pipeline — the rare case of a consumer video tool built for both a person and a script.

Worth watching this week: whether more desktop apps ship an official command-line surface for agents, or get one generated for them.

Which desktop app would you most want an agent to be able to drive? Write to us via our contact page.

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Open Source Radar — November 20: Magpie runs Claude Code on Kimi

Open Source Radar — November 20: Magpie runs Claude Code on Kimi

Today's open-source signal is in the seams around the agent, not the agent itself: which model it runs on, who investigates the build it just broke, and whether the harness wrapped around it can improve without touching the weights. Magpie (Go, 39,365 stars, MIT) is yetone's local gateway for a mess every agent user now lives in: Claude Code keeps its model in one file, Codex in another, each with its own key and base URL, and a subscription you pay for works in exactly one of them.

Open Source Radar — November 17: Google's Elephant-Goldfish Reviewer

Open Source Radar — November 17: Google's Elephant-Goldfish Reviewer

Today's board is agents doing the reading for you: a skill that counts what people actually posted, Google's clean-room reviewer for design docs, and a pipeline that turns lab notes into a paper. last30days-skill (Python, 68,470 stars) is an agent skill for the questions search engines are bad at. It pulls Reddit threads, X posts, YouTube and TikTok transcripts, Hacker News, Polymarket odds, GitHub activity and arXiv from the last 30 days, scores the pile by upvotes, likes and money at stake…

Open Source Radar — November 14: OpenAI's Codex Security CLI

Open Source Radar — November 14: OpenAI's Codex Security CLI

Today's board is all about review: OpenAI's vulnerability scanner now runs as an open CLI, Chrome is teaching coding agents the modern web, and a 10,000-star app that puts a human between the agent and the merge. codex-security (TypeScript, 13,345 stars) is OpenAI's scanner for finding, validating and fixing vulnerabilities in your own repository, released as an Apache-2.0 CLI and TypeScript SDK that started as a research preview on November 6 and gained 751 stars today alone.

Open Source Radar — November 13: Free Claude Code's 1.3B Tokens

Open Source Radar — November 13: Free Claude Code's 1.3B Tokens

Today's board is about routing around the bill: a proxy that points Claude Code at free models, a skills pack from Addy Osmani, Qwen's own coding agent, and a workspace where each bot gets its own computer. All four shipped code in the last 48 hours. free-claude-code (Python, 69,656 stars) is a local proxy that sits between your coding agent and whichever model provider you point it at — 12 harnesses and 61 providers, from Claude Code and Codex to Cline and Aider.

Open Source Radar — November 8: An Android Agent That Uses Your Apps

Open Source Radar — November 8: An Android Agent That Uses Your Apps

Today's board is agents on the move again — one that tapped into your phone's own apps, one that keeps coding after you close the terminal, one that reads a video frame by frame, and one built to make the first kind expensive to point at your binary. Mika (Kotlin, about 7,700 stars, Apache-2.0) puts the agent on the phone itself: wangzan101/Mika opens apps, taps, types and scrolls on a hidden background screen while you keep using the phone, finding buttons by looking at the screen rather…