Field notes from the M2 Max.
Local models, agents and products, tested on my own machine and written up in full. Every article here is the complete original text, with its media and interactive figures you can poke at.
Running a 2‑bit model as a coding agent: the Bonsai experiment
A 27B model at roughly two bits per weight, about 9 GB on disk, driven as a coding agent on an M2 Max. Two rounds, one app, and four hours of fighting a browser that was never there.
The fastest decode does not mean the best - Ornith 1.5 35B vs BigBang-v1 - Final Test
Same prompt, same M2 Max, same harness. BigBang finished in 54 minutes, Ornith took 119. Then I read the code.
Don't let the voxel image fool you. Ornith 1.5B 35B vs Bigbang-v1 voxel challenge
An underwater temple and a pagoda garden, two local models, and two screenshots that lie in opposite directions.
I've built an app using a 4bit MoE model. This is what I've learned
Five small local models, all 4-bit, one todo app prompt. What they built, what broke, and why the config mattered more than the model.
The complete beginner's guide to running local models in 2026 in plain English
If your Mac has 16GB or more of unified memory, you can run a real open model today. A complete walkthrough, no terminal required.
You don't need a mac to build SwiftUI iOS apps in 2026.
A native iOS SwiftUI app, built, run in the Simulator and tested by tapping, on a macOS devbox the agent got for the session.
The Ultimate Guide to Building Apps and Products in 2026 with Devin Local and Cloud
Desktop, cloud, CLI, Outposts, Fusion, DeepWiki, review, testing, security. Everything in one place, from someone who uses it every day.
Why mute, block, and 'Not interested' still leave slop in my X feed
I tried mute, block and Not interested. The slop kept coming, so I built the filter I wanted.
Devin took my macbook, tested my app and saved me hours of clicking!
My MacBook became Devin's machine for a day. Two Chrome profiles, one SMS code from me, and a full end-to-end report.
