Which AI brain fits your computer? Local models, explained plainly

Each brain has a plain name in the app, with the technical details one click away.

• Faster — Qwen3-4B Instruct (July 2025). Needs at least 8 GB of computer memory; download about 2.5 GB. Dense, 4 billion parameters, 4-bit.
• Balanced — Qwen3-14B (April 2025). Needs at least 16 GB of computer memory; download about 9 GB. Dense, 14 billion parameters, 4-bit.
• More capable — Qwen3-30B-A3B Instruct (July 2025). Needs at least 32 GB of computer memory; download about 18.6 GB. Mixture of experts, 30 billion stored / 3 billion active, 4-bit.
• More capable — high memory — Qwen3-30B-A3B Instruct, 8-bit copy. Needs at least 64 GB of computer memory; download about 32.5 GB. Same model as above at higher precision.
• Largest brain — OpenAI GPT-OSS 120B (August 2025). Needs at least 96 GB of computer memory; download about 63.4 GB. Mixture of experts, 117 billion stored / about 5 billion active.

Those are the actual model files, not nicknames for one model with five labels. The first three are different sizes of the same family; the fourth is a higher-precision copy of the third; the fifth is a different model entirely, from a different maker.

A model's parameters are the numbers it learned during training. More parameters generally means more room for knowledge and nuance, but also more memory and slower answers.

Size isn't everything. Newer, well-trained smaller models often beat older large ones, and a fast answer you can read now is often worth more than a slightly better one in two minutes. That's why we don't call any brain "the best" — only the largest, which is a fact.

The practical version: size sets a floor on what a brain can hold in mind and how much it knows, but the training it received matters just as much. Two models of the same size can feel completely different to talk to.

A dense model uses all of its parameters on every word it writes. Faster and Balanced are dense.

A mixture-of-experts model stores many specialized parts but only uses a few for each step. Qwen3-30B-A3B stores about 30 billion parameters and uses about 3 billion at a time; GPT-OSS 120B stores about 117 billion and uses about 5 billion. That lets them hold more knowledge while staying usable — but all the parts still have to fit in memory, which is why their requirements are higher.

The analogy that holds up: a dense model is one generalist reading your question cover to cover. A mixture of experts is a library where a few relevant shelves get pulled out for each question. The library knows more, but you still need a building big enough for the whole thing.

Model files are compressed so they fit on normal computers. A 4-bit copy is smaller and faster to load; an 8-bit copy keeps more precision but needs roughly twice the space. The "More capable — high memory" option is the same model as "More capable", just stored at 8-bit. It isn't a new generation and isn't guaranteed to answer better.

This compression is called quantization, and it's the reason a model that needed a server rack a few years ago now runs on a laptop. Every local model you use is a compressed copy of a bigger original.

The honest summary: 4-bit is the sensible default, and 8-bit is worth trying if you have the memory and want to compare answers yourself. It is not a different brain, and it should never be sold to you as one.

Sovereign reads your computer and greys out what won't run, so you can't make a bad pick by accident. But if you'd rather decide for yourself, here is the ladder.

• 8–15 GB of memory: Faster. Quick everyday questions and drafting.
• 16–31 GB: Balanced. Good for most people on most recent computers.
• 32–63 GB: More capable. Longer explanations, writing and structured thinking.
• 64–95 GB: You can try the high-memory copy of More capable.
• 96 GB or more: Largest brain, for demanding analysis on very large machines.

To check your memory on a Mac: Apple menu → About This Mac → look for Memory. On Windows: Start → Settings → System → About → Installed RAM.

You can switch brains at any time. Your conversations and Vault are kept; only the brain changes.

A local brain writes answers at a pace your hardware allows, so a smaller brain answers sooner and a large one makes you wait. On the same machine, the smallest brain can feel instant while the largest takes noticeably longer per answer.

That trade is the whole decision. If you find yourself waiting, a smaller brain that answers in a quarter of the time is usually the better daily companion. If you find yourself rewriting the question to get more depth, go up a size.

There is no score that settles this for everyone. The right test is one question you actually care about, asked to two brains, on your own machine.

Every model can only keep a limited amount of text in play at once — your question, the recent conversation, and whatever it pulls from your Vault. When a conversation runs long, that room fills up.

Sovereign manages this automatically so long conversations don't simply break: it keeps the parts that matter and trims the rest, and it tells you plainly when an answer was shortened rather than pretending nothing happened.

You also choose how long answers should be — Brief, Fuller or Detailed — in Response style. Longer isn't always better; a short answer that finishes beats a long one that stops halfway.

Sovereign doesn't charge per message and doesn't count your chat against a quota or your Online Vault space, because it all runs on your computer. The real limits are physical: how fast your computer is, and how much text a brain can hold in mind at once.

That's the quiet advantage of local AI for people who think out loud: the cost of an extra question is a few seconds of your computer's time, not a bill.

No. A brain never changes on its own, which means its answers stay consistent. When better open models appear, we add them as options, and you decide whether to switch. Your Vault and progress stay the same either way.

Consistency is underrated. If a brain answered a question a certain way in March, it will answer it the same way in September. Cloud services can and do change behavior under the same account without warning.

They're open-weight models published by their makers — the Qwen team and OpenAI — downloaded from a public model library and checked against a fingerprint so you know the file wasn't altered.

In practice: Sovereign downloads standard model files (GGUF format) and verifies each against a published SHA-256 fingerprint before using it. If the file doesn't match, it isn't used.

You can supply your own model file if you want to experiment. Sovereign marks those as unverified, because there is no fingerprint of ours to check them against — a fair warning rather than a blanket trust.

The engine that runs the files is built from the open-source llama.cpp project, pinned to a specific release and compiled and signed together with the app, so it doesn't quietly pull a different engine from somewhere else.

A brain reads your question, the recent conversation, and whatever it pulls from your Vault, then writes the answer a piece at a time. All of that has to fit inside the room it can hold in mind at once.

That's why the same brain can feel sharp on a short question and vaguer on a long one, and why a model with more training doesn't automatically keep more of your conversation in play. Knowledge and room are two different dials.

Sovereign sets those for you and manages them as a conversation grows. You see the effect rather than the settings: answers stay readable, long chats keep working, and when something had to be trimmed, the app tells you instead of going quiet mid-sentence.

Skip the benchmarks. Ask these four, on whichever brain your computer can run:

• Something you know cold — a fact from your own life or work. Did it get it right, and did it admit when your Vault didn't contain the answer?
• Something that needs following instructions: "answer in six bullet points, no more than eight words each."
• Something with a trap in it — a question that assumes something false. A good brain pushes back.
• Something long: "explain this to me in detail." Then check whether it actually finished, or stopped early.

Run the same four on a second brain and keep the one you prefer. Ten minutes tells you more about a model than any article, including this one.