Open Source Radar — November 3: GLM-5.2 on a laptop
Today's trending board has one theme: shrinking the machine AI needs. A C engine that streams a 744B model off your SSD, a tool that strips refusals out of a model in under an hour, desktop control that skips screenshots entirely, and a git wrapper built for running ten coding agents at once.
colibri (C, 46,865 stars) runs mixture-of-experts models far bigger than your RAM by keeping the always-on parts resident and reading the rest of the experts from disk as each token needs them. The colibri repo shows GLM-5.2 at 744B parameters in int4 with a 9.9 GB resident footprint and 32 seconds to ready on the CPU; the setup asks one question, measures your machine, and picks a model that fits 8 GB of RAM and 22 GB of disk. Version 2.1.0 shipped October 31 with thirteen engines, including Kimi K3, DeepSeek V4.1 Flash and Qwen3.8-Flash-Next, and it serves an OpenAI- and Anthropic-compatible endpoint on localhost. That last part is the interesting one: you can point a coding agent at your own laptop instead of an API bill. Disk reads cost latency, so this is a fit for batch work and long-context jobs, not chat snappiness.
Heretic (Python, 38,379 stars) modifies a model instead of just running it. It finds the directions inside the network that produce refusals and ablates them, optimizing against a refusal-rate scorer and a KL-divergence term so the model keeps its general ability while losing the "I can't help with that." The Heretic repo puts a Qwen3-4B pass at 30 to 40 minutes on an RTX 3090, and ships preset configs beyond the default censorship removal: less slop, less humor, less sycophancy, altered emotional tone. Version 2.0.0 landed November 1 with a plugin system so you can write your own scorers, and old 1.x configs break. Thousands of Heretic-processed models are already on Hugging Face, which tells you how much demand sat behind "abliteration" before anyone packaged it properly.
agent-desktop (Rust, 11,320 stars) is Vercel Labs' answer to computer use without the screenshot loop. Instead of guessing at pixels, it reads the OS accessibility tree and returns a JSON snapshot where every element an agent can act on has a ref; the agent clicks the ref, then snapshots again to confirm. The agent-desktop repo runs on macOS, Windows and Linux, exposes the whole thing as a CLI and an MCP server, and released version 1.1.0 this morning with Windows overlay and caret fixes. Accessibility trees are more reliable than vision for ordinary desktop software and useless for canvas apps and games, so treat it as the tool for spreadsheets, CRMs and internal admin panels rather than "an agent that plays your game." One caveat the README is honest about: snapshots and clipboard reads contain whatever was on screen, so keep those logs off public services.
Worktrunk (Rust, 10,208 stars) exists because git worktrees are the right answer to parallel agents and a miserable thing to type. One worktree per agent means Claude Code and Codex runs stop stepping on each other's changes, but plain git makes you name the branch three times to start one. The Worktrunk repo addresses worktrees by branch name, adds hooks on create and merge, generates commit messages from diffs, and shows CI status per branch in one list. The feature that actually matters at ten agents: on APFS, btrfs and XFS it shares ignored build directories copy-on-write, so ten worktrees get node_modules or a Rust target folder without building or copying ten times. Version 0.81 is current, installed via Homebrew, Cargo, Winget or pacman.
Worth watching this week: whether colibri's disk-streaming approach gets adopted by the agent tooling that already speaks OpenAI's API — local inference is only cheap if the latency hides behind a tool call.
Would you run a 744B model off your laptop's SSD, or is the disk-read latency a dealbreaker? Write to us via our contact page.
