Document how to run ExitBox in fully air-gapped environment

[olljanat]

Thanks for building ExitBox. It is definitely good starting point.
However, I'm very paranoid with these and would like to have fully air-gapped environment.

Afaiu, I would need at least:

• Fake opencode.ai in DNS and web server to host their binaries in there.
• Locally hosted LLM model.
• Server with powerful enough GPU to run all these.

There is a lot of articles in internet about local LLMs but I would like to get your honest opinion.
Which model(s) would be smart enough to be useful?

There is news that people are buying Mac Minis to run AI agents but maximum unified memory in those is 64 GB and I'm bit sceptic that it would be enough. With Mac Studio you get up to 512 GB but that is about 10k$ investment.

[thecodeassassin]

Hello! Excellent question and I was actually thinking about make a in-depth tutorial for just this!

I myself am running qwen3-coder-next and will be upgrading to qwen3.5 32b once it drops. You can run either with your macbook I believe and both are excellent.

You don't really need fake openai.ai because we vet the binaries and make sure it's installed from a secure source. I will think about your request but what we do is already many times more secure than raw-dogging Opencode.

I tried the following local models:

• Qwen3.5 397b (this runs on my GPU cluster, this you cannot run on your macbook) but this model is close to Opus 4.5 and ChatGPT 5.2
• Qwen3-coder-next (model I am currently using) - Fairly good results for the size

I tried a bunch more but these are the ones that give me very decent results. We are also training our own models but for very niche purposes (not AI code review).

I'll update this issue once the guide drops.

[olljanat]

• Qwen3-coder-next (model I am currently using) - Fairly good results for the size

Which quantization you are using? I know that it is way to reduce memory requirements but afaiu it is not free trade, if you go too far the model will become dope. That exactly my question that where the sweet spot is?

[thecodeassassin]

• Qwen3-coder-next (model I am currently using) - Fairly good results for the size

Which quantization you are using? I know that it is way to reduce memory requirements but afaiu it is not free trade, if you go too far the model will become dope. That exactly my question that where the sweet spot is?

I would always recommend at least Q4, Q5 is the sweet spot for quality if you have the RAM to run it with a decent context window size (32k+).

Q4 (4 bit) is a really good quant to run.

https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF

I am running this model with the Q5_K_M quant.

[olljanat]

btw. Have you considered supporting OpenClaw which would extend ExitBox also to non-coding related tasks?

I just found NetClaw which would be very useful but it definitely needs to be running air-gapped and have to work with read-only credentials to be safe enough to run in production environments.