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.







