A model page on this site is a structured copy of public fields: name, company, context window, parameter count, license, open-source flag, directory score, and sometimes a price or a benchmark. The paragraph under "About this model" is the vendor's or the card's description unless an editor's note sits underneath it and says otherwise. Read the fields before you read the paragraph. Marketing text is written to sound like every other flagship.
Name and company
The name is the identifier you will type into an API or a weights file. Alias rows that say "Latest" redirect to whatever the vendor currently calls the newest member of a family. They are convenient and they are unstable. Pin a dated name in anything you ship. The company line is who published the listing. An open weight rehosted by another organization may show a different company than the lab that trained it. Follow the source link before you assume you are talking to the lab.
Open source is a flag, not a freedom
The open-source badge means the listing was marked that way. It does not mean every weight file is under a license that allows your product. Read the license field, then read the license file upstream. Some weights are open for research and restricted for hosting a service. Open source and proprietary models walks through that choice. If the license field is empty, the page does not know, and you should not guess from the badge.
Parameters
The parameter count is in billions when we have a number. It is a size class. It predicts memory more reliably than it predicts whether the model can edit your repository. A mixture-of-experts model can quote a huge total and activate a smaller slice per token. The number on the page may be the total. Do not divide your GPU memory by that number and call it a plan. Use the VRAM field when it exists, and otherwise try the model.
Context window
The context window is how many tokens the listing claims the model can take in one call. A million-token window does not mean the model uses the middle of that window well. It means the vendor says the call will not be rejected for length at that cap. For a long repository, a long transcript, or a bundle of PDFs, a larger window is the field that decides whether the job fits at all. For a short classification prompt, the window is irrelevant and you should look at price instead. The comparison guide puts window, price, and license in that order on purpose.
Score
The directory score sorts the model list. It is not a benchmark average unless the benchmark block on the page says so. Two models a point apart are not meaningfully different because of that point. Use the score to decide what to open next, then ignore it. How ranking works says the same about skills and plugins.
Benchmarks, price, and VRAM
Benchmark bars are only as good as the names on them. A number without a dataset name is decoration. Compare two models on the same named benchmark, and prefer a task that resembles your work over a leaderboard you do not recognize.
Price, when present, is a snapshot of an input or output rate. Vendors change prices. Confirm on the vendor's pricing page the day you commit. A cheaper input price wins for bulk extraction. It loses when a single wrong answer costs more than the difference.
VRAM is the self-hosting field. If it is missing, the page does not know what card you need. Do not invent a requirement from the parameter count in your head and then blame the directory when the process is killed.
Modalities
Input modalities tell you whether the listing claims text, image, or other inputs. A text-only flagship will not describe a screenshot. That gap is why some inference plugins exist. If you need images and the model page says text only, switch models or add an adapter. Do not keep pasting images into a text endpoint and hoping.
A reading order
License, then context window, then whether you can host it, then price, then a benchmark that matches your task, then the prose. The prose is last because it rarely contains a number you can check. When you have two or three survivors, put them in the comparison tool and read the generated conclusion. It only talks about the rows you selected, and it will refuse to pretend a field exists when the page left it blank.
Walking a single page
Open a specific row, not an alias. A useful example in this directory is a named open model such as GLM-5.2 or Kimi-K3, or a named hosted model such as Claude Opus 5, rather than a card whose name ends in Latest. Read the company. Read whether the open-source flag is on. Read the context window and say out loud whether your largest real prompt fits. If the page lists about a million tokens and your prompt is a function, stop spending attention on the window. Look for a price or, if you will self-host, for VRAM and parameters. If those are blank, write "unknown" in your notes. Unknown is a fact the page gave you.
Then read the about paragraph and cross out every adjective you cannot tie to a field. "Flagship," "smartest," and "frontier" do not survive. "Text and image inputs" survives if the modality line agrees. "Redirects to the latest model" survives and means you should leave the page and open the target. An editor's note, if we wrote one, is the only paragraph that is allowed to say who should skip the model. Do not promote the vendor blurb into that role.
Two pages, one difference
Put the page you just read next to one alternative that differs in exactly one way you care about: open versus closed, or a much shorter window, or a listed price. Use the comparison tool so the numbers sit in one table, then follow deciding after a comparison. If you cannot name the one difference, you are collecting cards, not deciding. Close the extra tabs.
Bring the vendor's own pricing or license page up before you repeat a number in a design document. The directory is a cache. Caches lie when the upstream page moves, and the corrections guide is how you tell us the cache is wrong. Until it is fixed, your document should cite the vendor, not us. If you only have time to check one field against the vendor, check the one that would change your decision: the license if policy is the constraint, the price if volume is the constraint, the window if the prompt barely fits. Checking the adjective in the first sentence of the blurb is how an hour disappears without a decision. The rest of the directory, including the related-model grid at the bottom of the page, can wait until this row has a yes or a no. Related cards are a suggestion, not a tie you are obliged to break in the same sitting. The rest of the directory can wait. You opened this page to accept or reject a model, and the fields above are enough to do that when you are willing to write "unknown" and mean it.