Pick a model
Ask is model-agnostic: your agent can run on a proprietary frontier model or an open-source one. Pick deliberately based on quality, latency, and cost.
Prerequisites
- An agent draft, or an existing agent you can edit.
- An idea of what matters more for this agent: smartest answers, fastest answers, cheapest answers, or on-prem only.
Where the model picker lives
The model picker is hidden behind step 3 of the creator. Open your agent draft, switch to the Settings tab, scroll to Advanced settings, and click Configure next to Model configuration, labelled "Fine-tune the agent's core engine and reasoning settings." This opens the Agent model and settings dialog.
Inside the dialog
-
Left: the model list. A Search models box, a provider filter, and one row per model with a one-line description and its context-window size. Select a model to see its card on the right.
-
Right: model settings. Under Settings (with a Reset default):
- Creativity level: "Sets the Agent model's style. Choose Factual for precision or Creative for new ideas." Options: Deterministic, Creative, Factual, Balanced. The in-app tip: "Keep Creativity 'Balanced'. Lowering it can hurt its ability to reason."
- Processing limit: "Sets the maximum length of the response. Lower limits make the Agent model faster." Options Low / Mid / High plus a fine-grained Limit slider.
There is no raw temperature control or per-agent system-prompt field: Creativity level and Processing limit are the tuning surface. Click Save to apply.
Which models you'll see
Ask is model-agnostic, and the model catalog varies per deployment. Each client's instance is provisioned with its own set of proprietary and open-source models, and admins curate which of those appear in the picker via Settings → Capabilities → Model Settings (one model is marked Default for the org). Don't expect the list in this guide or in screenshots to match your instance; check the picker itself.
How to choose
| If you care most about… | Look at | Watch out for |
|---|---|---|
| Reasoning quality on hard tasks | The frontier tier of whichever proprietary families your instance carries | Higher per-token cost; slower TTFT |
| Latency / cost on routine tasks | Smaller variants in any family | May drop format constraints under load |
| Sovereign / on-prem deployment | The open-source models in your instance's catalog | Some features (e.g. native tool use) vary by model |
| Arabic-language performance | Check the picker, since some instances carry Arabic-tuned models | Tokenization quality varies by family |
Other engine settings worth knowing
The same Advanced settings block on the Settings tab also exposes:
- Context management: "Control how your conversation history is handled as it grows over time." Useful for long sessions where the agent should summarise or truncate older turns.
- Set safeguards: see Guardrails.
- Voice mode: toggle. "Allows you to speak to your agent and hear its responses." See Voice mode.
Tips & gotchas
Try the same prompt on two or three models before publishing: switch the model in the Agent model and settings dialog, then re-run the prompt in the Test panel. Differences in tool use, JSON adherence, and refusal style are often larger than headline benchmarks suggest.
Switching model mid-conversation can change behavior even with identical instructions. If you change the model on a published agent, re-test your most important prompts.