June 11, 2026
Offline AI, online chaos
Running Claude Code Offline on an M3 Pro with Qwen3.6
It finally worked on one laptop, but the comments turned into a full-on tech food fight
TLDR: A developer got a coding assistant to work entirely on a laptop with no data leaving the machine, proving private offline AI can do real work. The comments, though, were spicier than the setup: readers fought over the tool choice, the writing quality, and whether the slower model was worth it.
A developer says they got Claude Code—basically a coding assistant tool—to run fully offline on a MacBook, with nothing sent to the internet, and after a string of fixes it could investigate a software outage, write a patch, push a branch, and open a pull request all on its own. That’s the big flex: private, local, and actually useful. The catch? It was slow at first, then merely less slow, and the whole post is really about squeezing this setup into behaving on Apple hardware.
But the real entertainment is in the comments, where readers instantly split into camps. One crowd was basically: cool demo, but why use Claude Code at all? As tasuki put it, they get the appeal of keeping data on your own machine, but they’re baffled by the choice of tool, wondering if this is really the best option or just the one people are already locked into. Another camp came in swinging much harder, with corporealshift delivering the nuclear take that the article felt like “ai slop prose” and accusing it of being a recycled summary that “doesn’t respect the reader.” Ouch.
Then there’s the performance nerd side quest: InTheArena questioned why the author used this larger model at all, saying smaller options often seem faster in real-world reports. So the vibe is pure internet chaos: one part impressed by the offline magic trick, one part suspicious of the write-up, and one part arguing over whether the machine is smart or just taking the scenic route.
Key Points
- •The article presents a method for running Claude Code entirely offline by connecting it to a local Qwen3.6 model through Ollama on an Apple M3 Pro laptop.
- •It reports that an initial attempt timed out after ten minutes, but after four fixes the same machine completed incident investigation, patch creation, branch push, and pull request filing without data leaving the device.
- •The documented hardware and software stack includes Apple M3 Pro hardware, qwen3.6:35b-a3b-coding-nvfp4, Ollama 0.24.0, the MLX runner, and Claude Code v2.1.84.
- •The article states that the model's mixture-of-experts design allows a 35B-class model to run locally because only about 3B parameters are active per token.
- •It provides specific environment variables and launch commands for Ollama and Claude Code, including use of a local endpoint and omission of ANTHROPIC_API_KEY to prevent cloud access.