Start Now

> Model selector

The models

Ocai runs on open models, all served through the same interface. Two things separate them: how much they can hold at once, and how much thinking budget they will spend on a problem.

ModelContext windowThinking budget
Ocai Swift262KLow, High, Max
Ocai Core1MZero, Low, High, Max
Ocai Ultra1.3MLow, High, Max
GLM Pro1.3MLow, High, Max
GLM Flash1.3MLow, High, Max
DeepSeek Flash1MZero, Low, High, Max
DeepSeek Pro1MZero, High, Max

The picker lists the Ocai models first, then the rest in the order above. Every model starts on High.

Each family ships in two settings, named Pro and Flash. The Pro variant is the one to reach for on harder work, the Flash variant on faster work.

What a context window means

The context window is how much material a model can hold in one go: the conversation, your attachments, its instructions. Past that ceiling, the oldest material falls out of view.

This is the number to check before loading a lot at once. A folder of long documents, a codebase, a contract with everything attached. When the material is bigger than the window, Ocai works over it in pieces rather than pretending to hold all of it.

Long conversations are the other case. Once a chat outgrows the window, start a fresh one for the new subject.

Three models reach past a million tokens, which covers most real work in one pass.

Thinking budget

Thinking budget is how much a model works through a problem before it answers. The picker calls it a Thinking Budget, and it runs from Zero to Max.

  • Zero — answers directly. For a clear question with an obvious route.
  • Low — a little room to work things out.
  • High — problems with several parts, or where a wrong turn is expensive.
  • Max — the hardest work: multi-step analysis, awkward data, long chains of reasoning.

Raise the budget when a problem has moving parts. Leave it low when it does not, because the budget is spent whether or not the task needs it. See Agentic steps.

Not every model offers every level

The levels are not the same across the list, and the picker only shows what a model supports.

  • Ocai Core and DeepSeek Flash offer all four, Zero through Max.
  • DeepSeek Pro skips Low. Its levels are Zero, High, and Max.
  • Ocai Swift, Ocai Ultra, GLM Pro, and GLM Flash start at Low. Zero is not offered.

If you switch to a model that does not offer your current level, the budget falls back to that model's default, which is High across the board.

Changing the model

In a chat. The model picker sits with the composer. Open it, choose a model, then set the Thinking Budget from the flyout. The default for that model is marked, so you can see what you are overriding.

As your default. The model you pick becomes the default for your next new chat, and each conversation remembers the model you chose in it when you come back. Settings, then Preferences, holds the model used by sub-agents — worth pointing at a fast model, since sub-agents run in parallel and often on small pieces of work.

For images. Image models are chosen in the same place, and the generate button shows what the image costs in credits, so you can see the price before you generate.

The image model picker, with the credit cost shown on the generate button

Changing the model mid-conversation keeps the context, so you can start on one and finish on another.

Choosing in practice

  • Starting from nothing? Keep the default and raise the budget if an answer comes back shallow.
  • Loading a lot of material? Pick by context window, not by name.
  • Something important and hard? Raise the budget before switching models.
  • Repetitive work across many files? Keep the budget modest and let autonomous mode do the running.

Questions people ask

Does a bigger context window mean a better model?

No. It means the model can hold more at once. Quality comes from the model and the budget you set.

Why can I not pick Zero on some models?

Those models do not offer a Zero budget. The picker only shows the levels a model supports, which is why Ocai Swift, Ocai Ultra, and both GLM models start at Low.

Does a higher thinking budget cost more?

Yes, and it is the most direct way to control spend. Longer thinking means more output tokens. See Credits.

What are these models called internally?

The API names are the lowercase forms of the labels above, such as ocai_swift and deepseek_pro. Worth having when you are reading an error or contacting support.

Can I use a model I saw in an older version of Ocai?

The picker lists what is currently served. If you have a conversation pinned to a model that is no longer offered, it keeps working, and you can pick a current model when you are ready to change.