# Model Settings

> Per-model controls, shown exactly as each model accepts them - more than 40 parameters: native search, temperature, reasoning effort, verbosity, and which level wins.

- Canonical: https://big-agi.com/docs/model-settings
- Minimum tier: open  [Open ⊂ Free ⊂ Pro - a tier only ever ADDS to the Open baseline]
- Kind: reference · Last content update generation: 2026-07-27 · Review by: 2026-10-25
- Answers: What does temperature actually do? · What is reasoning effort, and is high worth it over medium? · Why can't I set temperature on this model? · I set it in two places - which one wins?

Every model carries its own controls, and Big-AGI shows exactly the ones that model accepts - more than 40 parameters in total, in every tier including **Open**. The full inventory: [big-agi.com/controls](/controls).

| Shortcut | What it does |
|---|---|
| `Ctrl + Shift + O` | `Current Model Options` - the options of the model in use |
| `Ctrl + Shift + M` | the Models dialog - change the options of any model |

Model settings live in the model options dialog, in the right-hand panel of the chat, and in a persona's model configuration. They layer: the persona overrides the model's global parameters, and the request overrides both.

## Native search

Native search is the preferred route to the web: the model finds pages and reads their content at the AI service. The major services run it natively.

[openai](/docs/connect-openai) · [anthropic](/docs/connect-anthropic) · [gemini](/docs/connect-gemini) · [xai](/docs/connect-xai) · [perplexity](/docs/connect-perplexity) · [moonshot-ai](/docs/connect-moonshot-ai)

Anthropic adds **dynamic filtering**: fewer tokens spent, more precise results. Forcing search on or off per message is the search control in the composer.



## The controls

| Control | Range | Notes |
|---|---|---|
| Temperature | `0.0 - 2.0` | unset: behaves as 0.5 |
| Effort | `low` · `medium` · `high` · `xhigh` · `max` | unset: high |
| Thinking Level | `minimal` · `low` · `medium` · `high` | unset: the model picks per prompt |
| Reasoning Effort | `none` · `minimal` · `low` · `medium` · `high` · `xhigh` · `max` | |
| Thinking | `none` · `low` · `high` · `max` | |
| Verbosity | `Detailed` · `Good` · `Brief` | the full editor spells `Good` as `Balanced` |
| Maximum Tokens | 256 to the model's own maximum, step 256 | unset: behaves as 8192 |
| Disable Streaming | on / off | unset: streaming stays on |
| vendor extras | cache breakpoints · code sandboxes · document skills · dynamic filtering · fast mode at a price multiplier · image generation options · a longer context window · thinking budgets | |

Those four names are one concept under four vendor labels; a model may offer only part of its range. Untouched, **Effort** already sits at high and **Thinking Level** lets the model pick per prompt, so on routine work the useful move is down.

| Setting | What it changes |
|---|---|
| **Temperature** | How far the model strays from its most likely wording. |
| **Effort** / **Thinking Level** / **Reasoning Effort** / **Thinking** | How much the model works before it answers. Raising it costs latency and thinking tokens. |
| **Verbosity** | Length and thoroughness of the reply, not its quality. |
| **Maximum Tokens** | The limit on one reply. |
| **Disable Streaming** | Waits for the whole reply instead of showing it as it arrives. |

The Temperature slider names itself as you drag:

| Temperature | The slider reads |
|---|---|
| `< 0.33` | `More strict` |
| `0.33 - 0.67` | `Creativity` |
| `> 0.67` | `Larger freedom` |
| `> 1` | `Extra hot ♨️` - past the top of the normal range |

## Which setting to change

| Symptom | Setting |
|---|---|
| The answer stops mid-sentence | **Maximum Tokens** |
| Repetitive, or three retries produce the same text | **Temperature** up |
| Invented specifics, or anything you will diff | **Temperature** down |
| Three paragraphs where you wanted one line | **Verbosity** `Brief` |
| You want the working on the record | **Verbosity** `Detailed` |
| Shallow on a problem with steps | the effort control up - or switch to a reasoning model |
| Slow and costly on routine work | the effort control down |
| Streamed replies arrive mangled | **Disable Streaming** |
| It keeps searching the web | the search control in the composer |

## Which level wins

| Level | What it holds |
|---|---|
| per-conversation override | composer, this thread only - the one you have is the search control |
| persona configuration | replaces your model settings outright, in that persona's chats |
| your model options | what you set, on that model |
| the model's factory value | shipped with the model entry |
| built-in fallback | 8192 maximum tokens, temperature 0.5 |

Top of that list beats everything below it. While an override is on, the panel shows the values it forces and locks the controls it touches.

Settings belong to the model you set them on: switch models and you get that model's own order, not a translation. `Ctrl + L` makes that switch mid-thread - the history continues, and the label above the chat changes to the model you picked. A persona that locked a model overrides the dropdown ([what a persona is](/docs/feature-personas)).

## Why a control is missing or greyed out

Four mechanisms, four fixes.

| What you see | Why | What to do |
|---|---|---|
| The control is not there at all | That model never declared it. Every extended setting is opt-in per model. | Nothing to fix - use a model that offers it. |
| The dropdown has fewer options than you remember | The model supports only part of its vendor's range. | Pick from what is offered. |
| Temperature is disabled and reads `Unsupported` | The model rejects temperature outright. | On OpenAI reasoning models, setting Reasoning Effort to `none` re-enables it. |
| Temperature greys out the moment you enable extended thinking | Extended thinking disables temperature at the vendor, so both surfaces disable it - including through an aggregator. | Turn thinking off to get temperature back. |

A control you have never touched reads `Default` - the model's own factory value is in force. The label changing is how you confirm your value took.

## Related

- [What a persona is](/docs/feature-personas)
- [Context windows, upload caps, and timeouts](/docs/context-and-limits)
- [Stalled or truncated answers](/docs/issue-messages)