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

# Models

> Find, compare, download, and manage local models.

## Choose a recommendation

**Discover** assesses your hardware and shows up to five recommendations. Start with **Balanced**, or choose Fastest, Faster, Smarter, or Smartest to shift the speed–intelligence trade-off.

## Read the model details

| Detail | How to use it |
| - | - |
| Intelligence | Compare model capability using its benchmark index. The displayed percentage is an index score, not an accuracy percentage. |
| Speed | Compare estimated generation throughput on your machine. It is an estimate, not a benchmark of your current task. |
| Memory | Check expected runtime memory, which is different from download size. |
| Fidelity / quantization | Lower-bit variants generally use less memory and disk space; higher fidelity preserves more of the original model. |
| Context | Check how much conversation and working material the serving configuration can accommodate. |
| Capabilities | Check whether the model supports features your task needs, such as vision or tool use. |
| Speculation | See the speculative-decoding method prepared for the model. Magnitude handles its setup. |

Use **Details** for source, license, and configuration information. Model licenses are separate from Magnitude's license.

## Browse Catalog

**Catalog** lets you search beyond the recommendations, filter by compatibility, and sort models. A model appearing in Catalog does not guarantee it fits your machine. Read its compatibility information before downloading.

## Download and manage models

Click **Download** for the selected model. **My Models** shows downloaded models and acquisition progress; its search, filters, and sorting help you find them later.

* **Load model** puts a downloaded model into memory. Replacing a running model asks for confirmation.
* **Stop model** unloads it without deleting the download. A later request can load it again.
* **Remove** deletes the installed model files to recover disk space.

Downloading and loading are different steps. A downloaded model uses disk space even while stopped; a loaded model also uses inference memory. Magnitude loads a model on demand when an agent requests it and unloads it when idle or memory is tight, so the first response after a while takes longer. Supported models use speculative decoding (MTP, DFlash, or DSpark) and prompt caching automatically.

Models are stored in `~/.magnitude/models`. To use another folder, see [Settings](/settings#model-storage).

## Change your agent's model

Choose a downloaded Magnitude model in your agent's model selector. If it isn't listed, reconnect the agent from **Connections**, then reload or restart it as described in its [integration guide](/integrations/overview).

For terminal control, see [CLI reference](/reference).

## Usage

**Usage** shows local inference activity by model for **Today** or **All time**.

| Metric | Meaning |
| - | - |
| Input tokens | Prompt and conversation tokens, including cached input. |
| Cached input | Tokens reused from cache; a subset of input tokens. |
| Output tokens | Generated tokens. |
| Generation speed | Output tokens per second. |
| Time to first token | Time before the response starts. Loading and prompt processing can delay it even when generation is fast. |


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.