Understanding models

What makes models different.

The words, numbers and scores you meet across the catalog, explained once, plainly.

The basics

Five ideas explain most of what separates one text model from another.

Tokens

Models read and write in tokens, pieces of words. In English a token is about three quarters of a word; French uses a few more per word. Prices for text are counted per million tokens.

This catalog turns that into something you can picture: credits per page written, a page being about 500 words.

Context window

Everything a model can hold in mind at once: your question, the conversation so far, the notes or files Sub Rosa hands it, and its own answer. Past the limit, the oldest part falls out.

128 thousand tokens is about 180 pages; one million is about 1,400. A bigger window helps with long documents, but it does not make a model smarter.

How much fits in a windowTokens, on the same scale
  1. A long note (10 pages)7,000
  2. A novel (about 300 pages)210,000
  3. 128K window128,000
  4. 1M window1,000,000

Reasoning

Some models think before they answer: they write a hidden draft, check it, then reply. It costs time and tokens, and pays off on math, planning, code and anything with several steps.

Many let you set the effort. For a quick rewrite, low effort is enough; for a hard problem, high effort is worth the wait.

Tool calling

The ability to ask Sub Rosa to do something instead of only writing text: search your notes, read a file, search the web, create an image. Sub Rosa’s agent is built on it.

A model without tool calling can chat, but the agent cannot work with it. Every family page says which versions have it.

Multimodal models

A model that reads more than text: photos, screenshots, scanned pages, sometimes audio or video. Sub Rosa switches to one when you attach a picture.

Under the hood

What the specification tables mean, and why two models of the same size can behave very differently.

Parameters

The numbers a model learned during training, counted in billions (B) or trillions (T). More parameters can hold more knowledge, but training quality matters as much as size.

Mixture of experts and active parameters

Many recent models are split into experts, and only a few experts work on each token. A model listed as 1T total and 32B active has the knowledge of a very large model and roughly the speed and cost of a 32B one.

All the experts hold knowledge; only the few chosen for each token do the work.

Open weights

The maker publishes the model itself, so anyone can run it on their own machines. That is why open models can be served in private mode by a provider that keeps nothing; closed models only run at their maker.

The license says what you may do with the weights. MIT and Apache 2.0 allow almost anything, including commercial use.

Private, anonymized, E2EE

Private: the provider that runs the model keeps nothing from your request once it has answered. Anonymized: your request reaches the model maker’s servers without anything that says who you are, and the maker’s own retention rules apply to its content.

E2EE versions are listed by their provider as encrypted end to end. That is the provider’s stated policy, not something Sub Rosa verifies, and these versions usually cannot call tools.

Credits and prices

One credit is one US cent. Text is priced per token, images per picture, video per second, voice per character, transcription per minute of audio. The app shows the exact price before anything is charged.

Reading a benchmark

Scores help you compare, as long as you know what they measure and who measured them.

What a benchmark tells you, and what it does not

A benchmark runs every model through the same test. It is the fairest way to compare, but it measures one thing on one day: a model that leads on code can trail at writing, and your own task is the real test.

Every score in this catalog shows who measured it and when. Small gaps rarely matter in practice.

Intelligence Index

Artificial Analysis runs about ten tests of reasoning, knowledge, math, code and agent work on every text model, the same way, and averages them into one score out of 100.

Two or three points apart is a close call. Ten points apart is a different class of model.

Arenas and Elo ratings

For images, video and voices there is no right answer to check, so arenas ask people. Two results of the same request are shown side by side without the model names, and the voter picks one. Thousands of duels give each model an Elo rating, as in chess.

Only the gap means something: 100 points apart, the higher model wins about 64% of duels. That is why the catalog shows Elo as dots on an axis rather than bars from zero.

What an Elo gap meansShare of duels the higher-rated model wins
  1. 25 points apart54 %
  2. 50 points apart57 %
  3. 100 points apart64 %
  4. 200 points apart76 %
  5. 300 points apart85 %

50% would be a coin toss.

Word error rate

For transcription: the share of words the model misses, adds or gets wrong. Lower is better. At 5%, one word in twenty needs fixing; under 3%, a transcript reads almost clean. Accents, noise and crosstalk raise it for every model.

Independent or announced by the maker

Independent scores come from a third party that tests every model the same way. Makers also publish their own numbers, with their own settings, at launch. The catalog keeps the two apart and labels the second kind.

Images, video and sound

The words that come up when you make pictures, clips, voices and music.

Text in images

Drawing readable letters is hard for image models: many still misspell signs and labels. Families that do it well say so, and it matters for posters, menus, packaging and slides.

Editing and retouching

An edit changes a picture from an instruction: the model may redraw much of it. A retouch changes only the zone you paint and keeps every other pixel. Sub Rosa offers both, in separate tabs.

Text, image or references to video

Text to video starts from a description. Image to video animates a picture you supply, which keeps its look. References to video takes several images of a character or product and keeps them consistent across clips.

Native sound and lip-sync

Recent video models generate the soundtrack with the picture: voices, effects, ambience. Lip-sync means the mouths match the words, sometimes in several languages. Without native sound, a clip is silent and you add sound afterwards.

Multi-shot and motion control

Multi-shot models cut between several shots inside one clip when the prompt describes them. Motion control copies the movement of a clip you supply onto your character.

Voices, lyrics and licensing

Expressive voice models follow tags such as [whispers] or [laughs] written in the text. Music models differ on lyrics: some require them, some refuse them. Models trained on licensed catalogs are the safer choice for anything you publish.