Slightly off topic, but Claude Code (and likely other models/harnesses) are incredibly effective at Nix. It can trivially self-verify, without side effects, which is a perfect match for an LLM.

If you've ever been put off by the difficulty of the language, it's worth checking it out again with AI assistance.

Absolutely. Running NixOS is a wonderful experience for this reason. NixOS + LLMs make it a breeze to make changes to your system, install and configure new software, and debug issues. Having every aspect of your system defined in a git repo is the perfect fit for agent harnesses.

I'll admit, even with LLMs to help, the Nix language and NixOS did have a rather steep learning curve, because it's quite different from anything I'd experienced before. But after getting the hang of it, I can't imagine going back to a "normal" OS and I'm very happy I put in the time to get over the initial friction.

I got curious one Saturday and ported our monorepo to use Nix for build and test and release. It was a very pleasant process. I had to put aside to tackle more pressing things but I'm all in on using Nix in that capacity.

Agreed, Claude has helped a lot with this project, and being able to iterate without side effects for system configuration changes is really a game changer for agents.

I also love how with the right system prompt they can pull in tooling for what they currently need via a nix shell

I've one-shotted custom distributions built with Nix using AI. It's crazy how well AI and Nix fit together.

There is also a microvm.nix project which helped us support sandboxes with firecracker. So, whole ai workflow pipeline can now be nixos.

Been running this on a few Asus GX10 machines with k3s on top, it’s been great. I’m running the new deepseek.

Thank you for your work!

What quant are you using, and what tps are you getting with K3?

The FP8 version from DeepSeek themselves [0], around 1800 tps prefill and 45 tokens per second decode.

I’ve been running a custom VLLM image with b12x as well as nvfp4_ds_mla.

I would say it’s quite fantastic in day to day, I use it mostly in Hermes and sometimes for coding.

I have qwen 3.6 27b on an rtx 6000 pro as well so I use that as a workhorse in pi with DS as a reviewer/planner.

[0] https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731

Edit: I think you may have misread my post. k3s is NOT kimi k3, and I did mention I was running deepseek.

Thanks - yeah, sorry, I mis-read that as you using both DS and K3...

Huge plus to anyone interested in the space to check out what Graham built here!

If anyone is also interested on Nix/CUDA/Capital Markets/Flox, we recently did another case study in the space - https://flox.dev/blog/deploying-hardened-flox-nvidia-cuda-st...

This has been amazingly helpful for managing my DGX Spark! Thank you for all your time and effort into this project!

Good to hear, thanks!

This is incredible. I have a Jetson lying around and will try to it out on this. I use it to play with vision models, not LLMs, and have been wanting a better way to manage the machine.

Thanks for sharing, saving this for when I get a DGX Spark