What people recommend, in their own words, with the receipt.
Co-founder of Ethereum; writes on cryptography, economics and public goods at vitalik.eth.limo.
10 things
piSoftware
The verdict: pi plus a basic searxng skill outperformed Local Deep Research. Also, pi is just much more configurable: I can easily just tell it to use not just internet searches, but also my own world_knowledge directory.
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
NixOSSoftware
The way that I use this daemon is that I run it on NixOS as a service, accepting requests on port 6000.
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
Arch LinuxSoftware Arch Linux
Software I have been a Linux user for a long time. About a year and a half ago I migrated over to Arch Linux.
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
llama-serverSoftware llama.cpp
As it turned out, ollama was not able to fit Qwen3.5:35B onto my GPU, but llama-server could. Hence, from that day forward, I resolved to cease being a cave-dwelling noob, and use llama-server (via llama-swap to make model swapping easier).
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
Qwen3.5:35BSoftware Alibaba Qwen
I have been using the Qwen3.5:35B model and have tried it on each of these, and I also tried the one-step-larger 122B.
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
BubblewrapSoftware
Sandboxing To keep my LLMs in check, I do most of my LLM usage from inside of a sandbox. I use bubblewrap for this.
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
SearXNGSoftware
I gave pi a skill for using the search engine SearXNG (which aggregates many search engines together at the same time), and one for calling into a daemon that I wrote that gives it access to read my email and Signal messages, and send-to-self, and send to others only with human confirmation.
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
OllamaSoftware
I used ollama before, but when I admitted to this in public half of Twitter told me that I was a noob and llama-server was clearly better and I must have been living in a very deep cave if I did not already know that. I tested their theory.
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
Local Deep ResearchSoftware
Its responses are, in my view, pretty bland and not very high-quality. I did a side-by-side test of asking Local Deep Research a question, then asking pi the same question (telling it to use searxng to make as many internet searches as needed), and I fed both outputs into an LLM to ask which is better.
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
NVIDIA DGX SparkHardware NVIDIA
I was not impressed with the DGX Spark; it's described as an "AI supercomputer on your desk" but in reality it has lower tokens/sec than a good laptop GPU - and on top of that, you have to figure out the networking details of how to connect to it from your actual work device etc.
His write-up of the local-first, privacy-hardened LLM setup he actually runs — hardware, models, agent tooling and sandboxing.
vitalik.eth.limo ↗·2026-04-02
Every entry is a verbatim quote and a link to the public post it came from. Nothing is paraphrased. If it is not on record somewhere public, it is not here.
The verdict is what they actually said: loved and liked are explicit; recommended means they told others to get it; uses means they only say they use it; read is a book on their public shelf with no verdict; mixed and disliked are kept too.
A dashed affiliate link, sponsored or their own label means the post disclosed a relationship. It is shown, never hidden.
The small bar beside a date shows how dated a recommendation is for its kind of thing.
The bar shows how dated a recommendation is, for its kind of thing. Full and green is recent. It drains, and turns amber and then grey, as the post gets older.
What counts as old depends on the thing. A laptop praised four years ago has been replaced twice; a book loved twenty years ago is probably still loved. The bar is empty after about:
A striped bar means the post carries no date and the date shown is an estimate. No date at all, no bar.