Keeping it local
Two years ago, opening a web browser with ChatGPT and getting it to write code felt like magic. These past few weeks I felt that magic again running a local language model on my Mac.
When I first used AI, the magic was that it could do a whole bunch of things. Now it can do those exact same things, but on my laptop.
The leaders, the frontier, are always going to be in a data centre, but the latest models like Qwen 3.6 27B, Gemma 4 31B, I can run myself.
They won't replace Opus 4.7 or GPT 5.5. But they feel like they're mine. The file is on my disk. The prompts don't leave my computer. There's no token meter ticking away limiting my usage.
The frontier will keep moving forward, but the trailing edge is getting exciting. Especially when you can run it from your phone.
A snapshot of how I'm using these models today.
- Claude Codedrives the work
- Opus 4.7frontier, the orchestrator I use most
- GPT 5.5frontier, second-opinion review
- Houtini-lmhands smaller jobs off to whichever local model is loaded
- Qwen 3.6 27Bthe heavy lifter27B · open weights · $0/tok · MacBook Pro
- Gemma 4 31Bmy daily driver31B · open weights · $0/tok · MacBook Pro
- Gemma 4 E4Bovernight jobs4B · 4-bit · via launchd · Mac mini