Blog

Ornith-1.0-9B is a small, tool-calling local model that's almost as good as Qwen3.6 35B A3B for a sovereign brain — but its 4-bit quant fits in ~5.6GB, so it runs on a 16GB machine instead of 32GB. Unlike Gemma 4, its skill and tool use actually work.

Gemma 4 12B looked great on the benchmark charts, so I tried it as the brain behind my sovereign brain. Real testing told a different story: 2x slower than Qwen3.6 35B A3B, half the context, and it tripped over its own filenames until it got stuck in loops.

I tested small local LLMs as the brain behind a sovereign second brain — Qwen3.6 35B A3B, Gemma 4 E4B, and LFM2.5 8B A1B — scored on reasoning, tool use, and whether they're actually a good thinking partner.

Localmaxxing — running your AI on your own hardware instead of the cloud. With NVIDIA's N1, AMD's Ryzen AI Max+, and DGX Spark all shipping 128GB unified memory, 2026 is the year it got practical. Here's the honest verdict on local vs cloud.

Can't run a local model? Point your sovereign brain at PPQ.ai's Trusted Execution Environment models — private cloud inference you pay for anonymously over Bitcoin Lightning with NWC auto-topups and Alby Hub.

How I run a private, local AI second brain on a laptop — llama.cpp, a local Qwen model, the Pi agent, and plain Markdown. 100% offline, no cloud, runs on 32GB RAM.