For Sovereign Inference

Choosing a model

Which model, which size, and what all those letters after the file name actually mean.

Open weightsRuns on your hardwareNothing phones home
A model is a file. Once it is on your disk it is yours, and it works whether or not anybody else is still in business.

Sovereign Inference runs open models in the .gguf format. You point it at a folder, it finds what is in there, and you pick one. No account, no queue, no rate limit, and no company deciding one morning that the model you built a workflow around is now deprecated.

The awkward part is choosing. A single model can be published in twenty versions with names like Q4_K_M and IQ4_XS, and almost nobody explains what separates them. This page does.

Two numbers decide nearly everything: how many parameters a model has, and how heavily it has been compressed. Get those right and the rest is taste.

Have one built

When none of them know your business.

Every model on this page is a general one. If you need something that knows your products, your documents, your policies and the way your customers actually ask about them, that is a different job, and it is one we do.

Custom AI Models
The first number

Parameters. Roughly, how much the model knows.

The B in 7B means billion parameters. More parameters generally means a better answer and a slower one, and it always means a bigger file. The memory figures below are rough, for the common Q4_K_M compression, and they are the number to check before you download 40 GB of anything.

1B to 4B~1 to 3 GB

Fast, small, and limited. Good for tidying text, quick summaries, autocomplete and anything you want instantly on a laptop with no graphics card. It will not reason its way through a hard problem, and it should not be trusted with facts.

7B to 9B~4 to 6 GB

The sensible default. This is where a local model starts feeling genuinely useful: drafting, rewriting, everyday questions, light code. Runs comfortably on most modern machines and on nearly any dedicated graphics card.

12B to 24B~8 to 15 GB

Noticeably sharper. Better at following long instructions, holding a thread and writing code that runs. This is the range where a 16 GB card earns its money, and where most people settle once they have the hardware for it.

27B to 34B~17 to 22 GB

Serious work. Real reasoning, long documents, proper coding assistance. Needs a 24 GB card to be quick about it, or a lot of patience and a lot of system memory.

70B and up~40 GB and up

Workstation territory. Excellent, and demanding. Either multiple cards, or a machine with a very large pool of fast unified memory. Running one of these off system RAM works, but expect to read the answer as it types.

Mixture of expertsReads big, runs light

The exception to the rule. An MoE model may have a huge total parameter count but only wakes a fraction of it per token, so it can run far faster than its size suggests. It still has to fit in memory in full, so budget for the whole file, not the active part.

The second number

Quantisation. How much precision was traded away to make it fit.

Models are trained at 16 bits per weight. Quantising rounds those weights down to fewer bits, which shrinks the file dramatically and costs some accuracy. The name on the file tells you exactly how it was done, once you know how to read it.

Q4_K_M

Four bits per weight · K style mixed precision · medium variant

Q

Quantised

It simply means the weights have been compressed below their original precision. Every file you download for local use will have it.

4

The bits

Roughly how many bits each weight is stored in. Lower is smaller and faster and less accurate. This is the single biggest lever on the file.

K

Mixed precision

K quants do not compress everything equally. The layers that suffer most from rounding keep more bits, the tolerant ones get fewer. Better quality for the same size.

M

S, M or L

How generous that mixing is. S is small and leanest, M is medium and the usual pick, L is large and keeps the sensitive parts richer still.

0

_0 and IQ

_0, as in Q8_0, is the older uniform method that predates K quants. IQ is newer and uses an importance map to decide which weights deserve the bits.

Q8_0 is the top of the ladder

There is no Q8_K_M or Q8_K_L in the standard set. At eight bits the loss is already so small that the clever mixing has nothing left to save, so Q8_0 is the practical maximum and is effectively indistinguishable from the original.

Some publishers add their own letters

You will see names like Q4_K_L or Q6_K_L from certain repackagers. Those usually mean the standard quant with the embedding and output layers kept at higher precision. Useful, slightly larger, and not part of the core naming.

The file size is the honest label

Whatever the letters say, the number of gigabytes tells you what you are actually getting. If two files claim different quality and weigh the same, they are the same trade wearing different names.

Which one to download

If you only remember one thing, remember Q4_K_M.

And remember the rule underneath it: a bigger model squeezed harder usually beats a smaller model kept pristine. A 13B at four bits will generally serve you better than a 7B at eight, for about the same memory.

Start here

Q4_K_M

The sweet spot, and the reason it is on almost every download page. Roughly a quarter of the original size for a quality drop most people never notice in normal use. If you are unsure, take this one.

If you have room

Q5_K_M and Q6_K

Measurably closer to the original for a moderate size increase. Worth it when the model is doing something exacting, such as code or long structured output, and you have the memory spare.

Near lossless

Q8_0

Essentially the full model. Rarely worth double the memory of Q4_K_M for conversational use, genuinely worth it for small models where the file is tiny anyway.

Tight on memory

IQ4_XS and IQ3

Importance matrix quants squeeze harder for the same quality, so an IQ4 often matches a Q4_K_M in slightly less space. They can be a little slower on some hardware. Worth trying when you are a gigabyte short.

Last resort

Q3 and IQ3

The damage starts becoming visible here. Instructions get missed, formatting slips, and long answers wander. Use it to get a model running at all, then move up as soon as you can.

Avoid

Q2 and below

At two bits a model is a shadow of itself. It still produces fluent sentences, which is exactly the problem, because the confidence survives the compression and the accuracy does not. A smaller model at Q4 is nearly always the better choice.

The families

Different houses are good at different things.

Every family below publishes open weights you can download and keep. We have deliberately not named specific versions or quoted benchmark scores, because both go stale within weeks and a page that pretends otherwise is lying to you. The character of each family holds far longer than its leaderboard position.

Qwen
The broadest range of sizes of anyone, from something that fits a phone up to workstation scale. Strong all rounder, particularly good at coding and at languages other than English. A sensible first download.
Llama
The most mature ecosystem there is. Enormous numbers of community fine tunes, and near universal tooling support, which makes it the safe default when something has to just work.
Mistral
Lean and quick. These models tend to punch above their parameter count and run well on modest cards, which makes them a good fit when speed matters more than the last few percent of quality.
Gemma
Built for modest hardware, with genuinely small options that still hold a conversation. Steady and predictable rather than adventurous, which is what you want from an always on helper.
DeepSeek
Reasoning and code. The distilled versions bring that ability down onto consumer hardware, while the full size models are firmly in serious workstation territory.
Phi and the small specialists
Small models trained on very carefully chosen data, which punch far above their weight on reasoning tasks while staying tiny. Narrower general knowledge in exchange.
Coder variants
Most families publish a code focused version. If you are wiring a model into development work, take the coder build rather than the general one. The difference is not subtle.
Embedding models
Not chat models at all. These turn text into numbers so documents can be searched by meaning, and they are what sits underneath any retrieval or document question setup. Small, fast, and a different job entirely.
The things that catch people out

Two numbers decide most of it. These decide the rest.

Memory, honestly

  • The file size is not the whole cost. The conversation itself occupies memory as it grows, and a long context can add gigabytes on top of the model.
  • Leave headroom. A model that exactly fills your card will slow to a crawl the moment the conversation gets long.
  • Sovereign Inference tells you before it fails. It flags when a model will not fit your GPU, CPU or RAM, rather than loading it and dying halfway through an answer.

Instruct, base and reasoning

  • Instruct or chat versions are the ones you want. They have been tuned to follow requests and hold a conversation.
  • Base models are raw completion engines. They are for building on, not for talking to, and they will frustrate you if you download one by mistake.
  • Reasoning models think visibly before answering. Far better on hard problems, considerably slower and more verbose on easy ones.

Licences and provenance

  • Open weights is not the same as open source. Some families are properly permissive, others carry conditions on commercial use. Check before you build a business on one.
  • Download from the publisher or a repackager you recognise. A model file is code that runs on your machine.
  • Once it is on your disk it stays working. No licence server, no deprecation notice, no morning where the thing you depend on simply stops.
What this page will not do

No leaderboard.
No best model of the month.

Rankings change faster than we could keep this page honest, and most published comparisons are measuring something other than whether the thing is useful to you. So we have told you how the sizes and the compressions actually work, and left the choosing to you. Download two, ask them both the same awkward question, and keep the one that answered it properly.