Sovereign Brain Ornith 1.0 9B: The Local Brain Bar Just Dropped to 16GB
For months the honest answer to “what should run my sovereign brain?” was Qwen3.6 35B A3B — genuinely smart, reliable with tools, a real thinking partner. The catch was always the same: it wants 32GB of RAM. That single number priced a lot of people out before they’d typed a word.
Ornith-1.0-9B changes the math. It’s a 9B model whose 4-bit quant fits in about 5.6GB, which means it runs comfortably on a 16GB machine — and it’s good enough to actually do the job.
A 9B that actually does the job
I put Ornith through the same gauntlet every candidate gets: be the engine behind my sovereign brain. The short version is that it works — with one honest caveat up front.
- It’s not the sharpest reasoner. It failed my carwash test — the silly little probe that sorts thinkers from parrots. Qwen3.6 35B passes it; Ornith doesn’t. On the hardest reasoning, the bigger model is still smarter, and I won’t pretend otherwise.
- But where it counts, it delivers. Tool calling is reliable. It calls the right tool when it should, and it doesn’t fall apart when a call comes back. This is the whole ballgame for a brain that has to read and edit your files.
- It reaches for its skills. It remembers it has structured workflows and uses them, so thoughts, todos, and decisions land in the right place automatically.
- It knows the brain. It understands which Markdown file to open and edit — no confusing
my_file.mdformy-file.md, no retry loops. - It’s a thinking partner. It analyzes, pushes back, and asks for the detail it’s missing, rather than just agreeing with you.
It also helps that Ornith is tuned for agentic, tool-heavy work in the first place — which is exactly the shape of the job a sovereign brain asks of it. Add a 256K context window and an MIT license, and there’s very little to argue with.
Unlike Gemma 4, the skill use is real
This is the part worth dwelling on, because we’ve been here before. Gemma 4 promised a small model that just works, and in real testing it didn’t: tool and skill use sort of worked — which in practice means it doesn’t, not reliably — and it got stuck in loops over its own filenames.
Ornith is the real deal where Gemma was “almost.” Tool calls and skill use aren’t a coin flip; they’re dependable. It’s not as seamless as Qwen3.6 — you’ll see the odd hiccup where the bigger model just glides — but it reliably reaches for the right tool and the right skill, which is the entire bar. This is the first sub-10B model I’ve run that clears it.
Almost Qwen, a quarter of the RAM
Let me be straight, because the house style here is honesty over hype: Qwen3.6 35B A3B is still the better model. It’s sharper on the hardest reasoning, and — because its MoE design activates only ~3B of its 35B per token — it’s actually a bit faster too, despite being far larger on paper. If you have the 32GB to run it, it stays my top pick.
But the gap is small, and Ornith gets close in roughly 5.6GB instead of needing a 32GB box. That’s not a marginal saving — it’s the difference between “I’d need to buy a machine” and “this runs on the laptop I already own.”
| Category | Score |
|---|---|
| Reasoning | ★★★☆☆ |
| Tool calling | ★★★★☆ |
| Skill use | ★★★★☆ |
| Knows the brain | ★★★★★ |
| Thinking partner | ★★★★☆ |
| Speed | ★★★☆☆ |
| Overall | 8 / 10 — recommended |
Which one should you run?
Pick by the RAM you actually have:
- 32GB or more: run Qwen3.6 35B A3B. It’s the better model — sharper reasoning and a bit faster — so if you have the capacity, use it.
- 16GB: run Ornith-1.0-9B (the
Q4_K_Mquant, ~5.6GB) to clear the bar on hardware you almost certainly already own.
Either way you get the thing that matters: a model that calls its tools, reaches for its skills, edits the right file, and pushes back when you need it to.
The barrier just keeps dropping
This is the localmaxxing story in miniature. The hardware on your desk didn’t change — the models got better and more efficient, and the floor to entry quietly fell with them. A year ago a private, local second brain meant a serious machine. Today it means 16GB and a free download.
Build your own
The tools are open source and free.
Take what you need and create your sovereign brain at sovereignbrain.me.