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

# Model Selection

> Selecting AI models, comparing multiple models, and tuning parameters

Cloosphere supports a variety of AI models. Pick the right one for your goal and tune parameters in detail.

## How to Select a Model

Choose a model from the dropdown in the chat header.

<Frame caption="Chat header model selector">
  <img src="https://mintcdn.com/cloocus/jo_Hh1nBizxj0BOc/images/en/chat/models-dropdown.png?fit=max&auto=format&n=jo_Hh1nBizxj0BOc&q=85&s=b48845b19812286b655290c6e3d34c02" alt="Chat header model selector" width="1912" height="906" data-path="images/en/chat/models-dropdown.png" />
</Frame>

<Steps>
  <Step title="Open the dropdown">
    Click the model selection area at the top of the chat.
  </Step>

  <Step title="Search and select a model">
    Pick a model from the list, or type a name in the search box to filter.
  </Step>

  <Step title="Set as default (optional)">
    Click **"Set as default"** below the model selector to use this model as the default for new chats.
  </Step>
</Steps>

<Tip>
  Type `@modelname` in the input box to ask a specific model directly, regardless of the currently selected model.
</Tip>

## Default Model

When you set a frequently used model as default, it's selected automatically every time you start a new chat.

<Steps>
  <Step title="Pick a model">
    Choose the model you want as default in the dropdown.
  </Step>

  <Step title="Save as default">
    Click **"Set as default"** below the model selector. A "Default model updated" notification appears.
  </Step>
</Steps>

<Note>
  The default model is per-user. If multiple models are selected when you save the default, the multi-model selection is saved as default.
</Note>

## Multi-model Conversations

Get answers from multiple models for the same question. Including the first one, you can select up to **4 models** simultaneously.

<Frame caption="Multi-model selection">
  <img src="https://mintcdn.com/cloocus/jo_Hh1nBizxj0BOc/images/en/chat/models-multi-select.png?fit=max&auto=format&n=jo_Hh1nBizxj0BOc&q=85&s=4fc9d99af79a080f84b5d563d513fddd" alt="Multi-model selection" width="837" height="481" data-path="images/en/chat/models-multi-select.png" />
</Frame>

<Steps>
  <Step title="Add a model">
    Click the **"+"** button to the right of the first model selector.
  </Step>

  <Step title="Pick additional models">
    Choose models to compare from the new dropdown. Up to 4 total (including the first).
  </Step>

  <Step title="Send the question">
    When you send the message, all selected models generate responses simultaneously.
  </Step>

  <Step title="Compare and merge responses">
    Compare the responses, and use **Merge Responses** to combine the best answers if needed.
  </Step>
</Steps>

<Frame caption="Comparing multi-model responses">
  <img src="https://mintcdn.com/cloocus/jo_Hh1nBizxj0BOc/images/en/chat/models-multi-response.png?fit=max&auto=format&n=jo_Hh1nBizxj0BOc&q=85&s=8ed7147e22029c4fd4c4b280321228fe" alt="Comparing multi-model responses" width="1911" height="907" data-path="images/en/chat/models-multi-response.png" />
</Frame>

<Accordion title="Multi-model use cases">
  * **Quality comparison**: Compare model performance with the same question
  * **Cross-validation**: Get multiple opinions before important decisions
  * **Best-answer selection**: Combine each model's strengths into the best answer
  * **Cost vs. performance**: Compare results from low-cost and high-performance models
</Accordion>

<Note>
  The multi-model feature can be restricted for regular users by admin settings. If you don't have permission, the "+" button isn't shown.
</Note>

## Tuning Model Parameters

The **Chat Controls** panel lets you tune model parameters per conversation.

### Opening the Panel

| Environment | Path                                                                                   |
| ----------- | -------------------------------------------------------------------------------------- |
| **Desktop** | Top-right **More (⋯) menu** in the chat → select `Overview` or `Artifacts`             |
| **Mobile**  | Tap the **slider icon** on the right of the header, or open the More menu → `Controls` |

<Note>
  The desktop More (⋯) menu only appears **after sending the first message**.
  In a fresh chat, send a prompt once to start the conversation, then tune parameters.
  (Adjusted values apply to subsequent messages.)
</Note>

Opening the panel with **Overview** first shows the conversation flow diagram. Click the **← (back)** arrow at the top left of the panel to switch to the **Chat Controls** view, where you can adjust the system prompt and advanced parameters.

<Frame caption="Conversation flow diagram in the Overview view (use the ← arrow at the top to switch to Chat Controls)">
  <img src="https://mintcdn.com/cloocus/jo_Hh1nBizxj0BOc/images/en/chat/models-controls-overview.png?fit=max&auto=format&n=jo_Hh1nBizxj0BOc&q=85&s=97eb1371904706b0cbe94873784e1602" alt="Conversation flow diagram in the Overview view" width="1912" height="907" data-path="images/en/chat/models-controls-overview.png" />
</Frame>

<Frame caption="Chat Controls panel">
  <img src="https://mintcdn.com/cloocus/jo_Hh1nBizxj0BOc/images/en/chat/models-chat-controls.png?fit=max&auto=format&n=jo_Hh1nBizxj0BOc&q=85&s=c3bfe5a1c09bd0fead4e17757413954f" alt="Chat Controls panel" width="619" height="907" data-path="images/en/chat/models-chat-controls.png" />
</Frame>

### System Prompt

Set a system prompt that applies to the entire conversation.

```
Example: "You are a Python expert. All code must include type hints
and follow PEP 8 style."
```

### Key Parameters

<Accordion title="Seed">
  Setting a fixed seed produces (nearly) identical responses for the same prompt.
  Useful when you need reproducible results.
</Accordion>

<Accordion title="Stop Sequence">
  When the specified string appears, the model stops generating.
  Use it to control specific output formats.
</Accordion>

<Accordion title="Temperature">
  Controls response creativity (randomness).

  | Value          | Behavior                          | Use Case                                         |
  | -------------- | --------------------------------- | ------------------------------------------------ |
  | **0.0 \~ 0.3** | Deterministic, consistent answers | Code generation, fact-checking, data extraction  |
  | **0.4 \~ 0.7** | Balanced answers                  | General chat, summarization, translation         |
  | **0.8 \~ 2.0** | Creative, varied answers          | Brainstorming, creative writing, idea generation |

  Range: `0.0` \~ `2.0` (UI slider initial value: 0.8 — the value shown when switching to custom; this may differ from the model's actual default)
</Accordion>

<Accordion title="Reasoning Effort">
  Applies only to reasoning models. Controls how much effort goes into reasoning.

  * Use values like `low`, `medium`, `high`.
  * Not supported by all models.
</Accordion>

<Accordion title="Top P">
  Limits the probability range the model considers when picking the next token.

  * **0.1**: Considers only the top 10% probability tokens — very focused
  * **0.9**: Considers the top 90% — diverse expression
  * Generally used together with Temperature; we recommend tuning only one at a time.
</Accordion>

<Accordion title="Frequency Penalty">
  Discourages reuse of tokens that already appeared. Higher values reduce repetition.

  Range: `-2.0` \~ `2.0` (default: 0)
</Accordion>

<Accordion title="Max Tokens (num_predict)">
  Limits the maximum length of the AI response in tokens. If unset, the model's default max applies.

  <Warning>
    Setting it too low can cause responses to be cut off mid-stream.
  </Warning>
</Accordion>

### Other Advanced Parameters

| Parameter                                         | Description                                                                                       |
| ------------------------------------------------- | ------------------------------------------------------------------------------------------------- |
| **Stream Chat Response**                          | Real-time streaming response on/off                                                               |
| **Function Calling**                              | Tool calling mode (Default / Native)                                                              |
| **Logit Bias**                                    | Direct adjustment of the appearance probability of specific tokens (-100 \~ 100)                  |
| **Mirostat**                                      | Algorithm that auto-regulates response perplexity                                                 |
| **Mirostat Eta**                                  | Learning rate of the Mirostat algorithm — controls how quickly feedback is applied                |
| **Mirostat Tau**                                  | Mirostat's target perplexity — balances output coherence and diversity                            |
| **Top K**                                         | Considers only the top K tokens for sampling                                                      |
| **Min P**                                         | Considers only tokens with at least the minimum probability                                       |
| **Presence Penalty**                              | Applies a flat penalty to already-seen tokens to encourage new topics (range: `-2.0` \~ `2.0`)    |
| **Repeat Last N**                                 | Number of recent tokens checked for repetition                                                    |
| **Tfs Z**                                         | Tail-Free Sampling strength — removes low-probability (tail) candidate tokens (`1.0` disables it) |
| **Tokens To Keep On Context Refresh (num\_keep)** | Number of tokens to retain when the context window is refreshed                                   |
| **Repeat Penalty (Ollama)**                       | Penalty on repeated tokens (Ollama-only, range: `0.0` \~ `2.0`)                                   |
| **Context Length (Ollama)**                       | Length of the context the model references (in tokens)                                            |
| **Batch Size (num\_batch)**                       | Number of tokens processed in a single batch                                                      |
| **use\_mmap (Ollama)**                            | Whether to load the model file via memory mapping (mmap)                                          |
| **use\_mlock (Ollama)**                           | Locks the model in memory (mlock) to prevent swapping                                             |
| **num\_thread (Ollama)**                          | Number of CPU threads used for inference                                                          |
| **num\_gpu (Ollama)**                             | Number of layers to offload to the GPU                                                            |

<Note>
  The Chat Controls panel itself is visible to all users, but the **System Prompt** and **Advanced Params** sections are only shown to admins or users with the chat-controls permission (`permissions.chat.controls`).
</Note>

## Model Capability Display

Agent models manage each capability (web search, image generation, code execution) individually.

| State    | Description                            |
| -------- | -------------------------------------- |
| **on**   | Enabled by default                     |
| **user** | User can enable manually (default off) |
| **off**  | Disabled (not shown in the input menu) |

For base models (those without `base_model_id`), all capabilities enabled by the admin in system settings are shown.