The model directory splits listings into open source and closed source, with multimodal, code, and reasoning cutting across both. The flag is a starting filter. The decision is about who runs the weights, who may see the prompt, and which license you can sign.
What the flag is claiming
An open-source mark on this site means the listing was published as open weights or open code, not that every downstream file is free for a commercial API. A closed-source mark means you are expected to call someone else's service. Alias rows that point at "latest" are almost always closed services, because the vendor can move the pointer. Pin a specific name either way.
Open the license field and the source link. If you are about to ship a product, the license file is the document, and this page is the index that helped you find it. Our terms say the same thing: we do not grant you rights in somebody else's weights.
Choose open weights when the constraint is control
Pick an open listing when at least one of these is true:
- The prompt contains source code, customer text, or anything your policy says must not leave a network you control.
- You need to freeze a build. A hosted model can change behavior on a Tuesday. A weight file you hashed does not.
- You need to negotiate, audit, or fine-tune. A proprietary API may forbid training on outputs or may simply not offer weights.
- The bill for a high-volume internal task would exceed the cost of a machine you already have.
The cost of that choice is operations. You need a card with enough memory, a runtime, and a person who notices when the process falls over. The VRAM field, when the page has it, is the first sizing hint. Kimi-K3 and GLM-5.2 are examples of open listings with long context windows in this directory. Search those names from the model index if the link on a card moves. Read the editor's note on the page when we have written one. The parameter count alone will not tell you whether your GPU survives the load.
Choose a proprietary API when the constraint is time
Pick a closed listing when at least one of these is true:
- You need a result today and you do not want to operate a server.
- The vendor's current model is clearly ahead on the task you measured, and the data you send is allowed to leave.
- You want a support channel and a pricing page more than you want a weight file.
- The workload is bursty. Paying per token beats owning a GPU that sits idle.
Claude Opus 5, the Qwen Max listing, Gemini Flash, and Grok are the kind of closed or hosted names this directory sorts near the top. "Kind of" matters: some Qwen and DeepSeek rows are open weights and some are hosted endpoints with the same brand. Read the company line and the open-source flag on the specific row. Two cards with similar names can be different products.
Multimodal, code, and reasoning are not a license
A multimodal model accepts more than text, usually images. That says nothing about the license. A code model is aimed at programming tasks. It can still be the wrong size for your repository. A reasoning model spends more tokens thinking before it answers. That can help on hard bugs and can waste money on "extract the email address." Filter by those categories after you have decided whether the weights may live on your machine.
A comparison that does not cheat
Put one open model and one closed model that both claim to fit the job into the comparison tool. The written conclusion will name the context windows and the license split if the fields exist. If a field is blank, the conclusion will not invent it. Then run the same prompt against both, with data you are allowed to send to the closed API. Keep the one that finishes the real task. A directory score a few points higher is not a reason to switch a production setting.
If the open model only wins because you refused to send the data out, that is a valid win. Write that constraint down next to the model name so the next person does not "upgrade" you onto a hosted alias.
A decision you can hand to a teammate
Write five lines and stop. The data class, either internal source or a public sample. The model name you pinned, not an alias. Whether it is open or hosted, and the license name if you found one. Where the process runs, a machine you control or a vendor account. The prompt you tried, and whether the answer was acceptable.
Those lines are the decision. A screenshot of our comparison table is not. The table will change when scores change, and your teammate cannot reconstruct the policy from a highlighted cell. If the data is internal and the model is hosted, the decision is inconsistent and you are not done. Either change the data you send or change the model. How to read a listing is how you check that the row you named still says what you think it says.
Teams get into trouble when one person picks an open model for privacy and another person later calls a same-brand hosted endpoint because the name looked familiar. DeepSeek, Qwen, and similar families have both kinds of rows in this directory. The brand is not the flag. Say the flag every time you write the model down.
If you are choosing for a hobby project with public data and no budget for a GPU, a hosted model with a small quota is a reasonable place to run it. Write that down too, so you do not later describe the hobby setup as an architecture. The first week guide assumes this kind of honesty: one pinned name, and a sentence about what the agent is not allowed to do. A directory score is not one of the five lines. Leave it out so nobody treats a sort key as a requirement. If the five lines took you longer than the comparison, that is the decision working. The table was only a way to see the fields. The lines are what you will still agree with next week. Re-read the five lines the next morning before you change a default in a shared environment. If you cannot explain line 4, where it runs, you picked a name and not a setup. Go back to the listing and finish the hosting question before anyone else copies your choice. A teammate who copies only the model name will copy the brand and miss the flag, which is the failure this section exists to prevent. Write the flag on the same line as the name. The comparison guide is the same discipline with two columns in front of you. Use it when the five lines are done for each candidate and you still have more than one name left.