> ## Documentation Index
> Fetch the complete documentation index at: https://docs.callkaro.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# LLM Model

The model panel picks the brain that writes the replies, the backup brain, and how much work the agent may do before it has to answer.

## Primary Model

The model every reply goes through. Available models fall into two kinds, and the panel changes depending on which you pick.

| Kind                   | Models                                                                                                        | Extra control    |
| ---------------------- | ------------------------------------------------------------------------------------------------------------- | ---------------- |
| **Temperature models** | `callkaro/arjuna-2.5`, `gpt-4.1`, `gpt-4.1-mini`, `gpt-4.1-nano`, `gpt-4o`, `gpt-4o-mini`, `gemini-3.5-flash` | Temperature      |
| **Reasoning models**   | `o4-mini`, `gpt-5-mini`, `gpt-5-nano`, `gpt-5.4-mini`, `gpt-5.4-nano`, `gpt-5.1-chat`, `gpt-5.2-chat`         | Reasoning Effort |

<Tip>
  **Start on `callkaro/arjuna-2.5`.** It is CallKaro's own model and the default for new chat agents. Move to a larger model only when you have a concrete failure it does not handle.
</Tip>

### Temperature

`0` to `2`, default **0.5**. Low is precise and repeatable; high is more varied. For support and sales agents that must stay on-message, stay at or below 0.7.

### Reasoning Effort

`low` / `medium` / `high` (and `xhigh` on `gpt-5.2-chat`), default **medium**. Higher effort handles more complex logic but is slower and costs more — and on a chat channel, slower is visible to the customer as a longer wait.

## Secondary Model

A required backup, defaulting to `gpt-5.4-nano`.

<Note>
  The secondary is used **only when the primary errors** — a timeout, a 5xx, or a rate limit. It is not a quality fallback: a primary that replies badly is still the reply the customer gets.
</Note>

Pick a secondary from a **different provider** than the primary. A backup on the same provider does nothing during a provider-wide outage, which is the case it exists for.

## Max Function Rounds

`0` to `6`, default **3**.

How many rounds the agent may spend calling functions before it must reply. The reply itself is always one additional round.

| Value           | Behaviour                                                                    |
| --------------- | ---------------------------------------------------------------------------- |
| **0**           | No function-gathering pass. If the agent has functions, one round still runs |
| **3** (default) | Enough for a lookup, a follow-up lookup, and a send                          |
| **6**           | For chains where one function's output feeds the next several times          |

Each round is another model call, so higher values mean slower replies and higher cost. Raise it only if you can see the agent running out of rounds mid-task.

## Related

<CardGroup cols={2}>
  <Card title="Conversation & Features" icon="sliders" href="/chat-agents/configuration/conversation-and-features">
    History, limits and channel features
  </Card>

  <Card title="Functions" icon="code" href="/chat-agents/functions/overview">
    What those rounds are spent on
  </Card>
</CardGroup>