# Presets

Save settings in [Settings → Presets](https://zalmir.ai/account/settings#presets), then reference them in `POST /api/v1/chat/completions`. Every API key on your account uses the same presets. Other accounts cannot use them.

## Create and manage

Choose **Create preset**, enter a slug, a model or fallback models, and an optional system prompt. Save, edit, or delete from the same section. Slugs use lowercase ASCII letters, digits, and hyphens, with 1–48 characters. You can keep up to 50 presets; a system prompt can contain up to 32 KB of UTF-8 text.

There is one current configuration per slug, with its last-updated time. Saving the same slug replaces it; the last save wins. Presets are private account settings, not organization-shared configurations. There is no version history, rollback, or preset-management API.

## Reference a preset

```json
{
  "model": "@preset/email-copywriter",
  "messages": [
    {
      "role": "user",
      "content": "Write a short welcome email."
    }
  ]
}
```

You can also send a concrete model and a `preset` field:

```json
{
  "model": "deepseek/deepseek-v4.1-flash-abliterated",
  "preset": "@preset/email-copywriter",
  "temperature": 0.7,
  "messages": [
    {
      "role": "user",
      "content": "Write a short welcome email."
    }
  ]
}
```

The combined form `model: "deepseek/deepseek-v4.1-flash-abliterated@preset/email-copywriter"` works too. The `preset` field accepts `@preset/email-copywriter` or the bare slug `email-copywriter`. Prefer the prefixed form; bare slugs work only on Zalmir. If both fields name a preset, they must name the same one.

## Request fields take priority

- Defaults and the request are shallow-merged: a supplied request field replaces the matching preset field. Fields absent from the request retain their saved defaults. Arrays are replaced, not appended.
- A concrete request model, including the model before a combined reference, overrides the saved model. A direct `@preset/slug` reference only selects the preset; it does not replace its model.
- A request `system` or `developer` message overrides the saved system prompt. Otherwise the saved prompt is prepended as a system message. Other request messages are unchanged.
- An inherited `models` list remains unless the request supplies its own `models`. The explicit model is tried first. If neither the preset nor the request supplies a model, existing account defaults apply.

Fallback lists accept 1–3 model IDs and use the gateway's existing fallback rules: your account must allow fallbacks, request restrictions still apply, and at most three attempts are made. Saving a preset does not make an unavailable model available or bypass privacy rules, spending limits, request bounds, or billing.

**Which settings a preset request uses**

1. **Request** Fields you send win over the preset.
2. **Preset** Fields you leave out keep the preset's saved values.
3. **Account** No model in either? Your account defaults apply.

## Stored generation defaults

The JSON parameters editor accepts only `temperature` (0–2), `top_p` (0–1), `seed` (a safe integer), `presence_penalty` and `frequency_penalty` (−2–2). These fields are forwarded to the provider; their effects have not been verified on the current models.

`max_tokens` (a positive integer) may also be stored as the **output limit and billing reservation**. Answers stop at about that limit, streaming or not; a few extra tokens may arrive, and you are never charged for more output than the limit. This is not exact token counting. The normal request validation and tariff reservation limits still apply. If you send `max_completion_tokens`, remove or explicitly override an incompatible saved `max_tokens`; differently named fields are not aliases during the shallow merge.

Presets do not store stop sequences, output-format controls, tools, plugins, or provider-routing rules. They are not available in website Chat.

## Updates and errors

An unknown preset, or one that belongs to a different account, returns HTTP 400 `preset_not_found` before a billing hold. Malformed or conflicting references return HTTP 400 `invalid_preset`; unavailable preset storage returns HTTP 503 `presets_unavailable`. No inference is started for these refusals.

Each call reads the current saved configuration once. Its effective merged body is used for validation, request digests, and billing. If a preset changes between two calls with the same `Idempotency-Key` and the effective body changes, the second call conflicts instead of silently using the old configuration.

## Stored configuration

Your saved system prompts and settings are stored on Zalmir's servers, readable by the service, until you delete the preset or close your account. This is not end-to-end encryption. Preset content is excluded from application logs and masked in session recordings. API conversation messages and answers are not saved as part of a preset. See the [Privacy Policy](https://zalmir.ai/privacy#content).

Source page: https://zalmir.ai/docs/guides/presets
