DEV Community

Cover image for Seven Hours, Zero Internet, and Local AI Coding at 40,000 Feet

Seven Hours, Zero Internet, and Local AI Coding at 40,000 Feet

Sebastien Castiel on September 03, 2025

Seven hours on a plane, zero internet, and a curiosity about local AI coding. That's how I found myself downloading 13GB of AI models in an airport...
Collapse
 
jess profile image
Jess Lee

This was a fun read, thanks for sharing!

Collapse
 
etienneburdet profile image
Etienne Burdet

In my latest experiments, I used devstrall through Ollama and then accessed via KiloCode. It was easy enough to setup.

But yeah, totally agree. I feels "almost there". Kimi dev is pretty awesome too, although bigger than devstral.

Collapse
 
umang_suthar_9bad6f345a8a profile image
Umang Suthar

The biggest insight for me is how local AI coding feels both liberating and limiting at the same time, you get independence from APIs and quotas, but pay the price in speed, heat, and battery life. It feels like the story of where we are with AI right now: powerful enough to glimpse the future, but still early enough to remind us how much room there is to grow.

Collapse
 
john4650hub profile image
JohnDelvin

I actually did something similar. My approach differ from yours slightly because it uses internet to access the app's front end hosted locally in a github workflow run and exposed to the internet using ngrok and also i don't get the performance issues you got on your physical machine. All i do is persist the workflow run(runs under a public repo) for about 6hours. Iam implementing a persistent storage using mega so that the important stuff is retained.

Collapse
 
rajatkantinandi profile image
Rajat Kanti Nandi • Edited

Good to know the current state of it from your article and realizing that it is not ready for adaptation yet. I think local AI might not be ready for full vibe coding but it can definitely do code completion part to some extent on a relatively older device as well.
I am used to cursor code completion and now it feels odd without using some form of code completion. But it does not work offline.
So, one weekend when I was in a no network zone, I decided to try out XCode’s offline code completion and I was surprised that it actually worked. Although the quality is not nearly as close as cursor but it does a very decent job. I have been building an iOS app and I was mostly using cursor to code and XCode to only compile the code. But I had downloaded Apple’s code completion model anyway. Haven’t thought I would use that but it kind of works!

I am not sure if more companies will offer offline code completion models in their IDEs but I think they should. Although it is probably not in their best interest I guess.

BTW, I used macbook pro M1 with 8 GB RAM for this purpose.

Collapse
 
igor_k_151f8c29ae911834ba profile image
Igor K

Just curious - if buy a Framework Desktop which costs 2K$:

  • 16-core/32-thread CPU, up to 5.1GHz
  • 40 Graphics Cores, up to 2.9GHz
  • 128GB of non-upgradeable LPDDR5x

will it work smoothly here and compensate all the investments (instead of paying openai subscription)?

Collapse
 
thebrown profile image
Saleumsack

Thanks for sharing!

Collapse
 
avinashzala profile image
Avinash Zala

Seven hours on a flight usually feels endless, but turning it into an AI coding lab sounds like the ultimate productivity hack

Collapse
 
naveen_kumar_17 profile image
Naveen Kumar

This gave me a interest to run a local model on my device like this