# Models

> Registry IDs, model files, downloads, verification, and removal.

Canonical: https://loxa.dev/docs/models

Loxa stores managed model files in `~/.loxa/models`. This page describes public
`origin/main@a59aec2`.

## Registry [#registry]

The built-in registry contains these six IDs:

| ID                       | Params  | Quant    | License label |
| ------------------------ | ------- | -------- | ------------- |
| `qwen3-coder-30b-a3b-q4` | 30B-A3B | Q4\_K\_M | apache-2.0    |
| `qwen3-coder-30b-a3b-q8` | 30B-A3B | Q8\_0    | apache-2.0    |
| `qwen25-coder-7b-q4`     | 7B      | Q4\_K\_M | apache-2.0    |
| `qwen25-coder-7b-q8`     | 7B      | Q8\_0    | apache-2.0    |
| `gemma-3-4b-it-q4`       | 4B      | Q4\_K\_M | gemma         |
| `qwen3-14b-q4`           | 14B     | Q4\_K\_M | apache-2.0    |

The model license labels are registry metadata for the model artifacts. They are
separate from the Loxa project's [Apache License 2.0](/docs/project#license).

## Files and partial downloads [#files-and-partial-downloads]

Each registry entry has a final GGUF filename. While a transfer is incomplete, Loxa
writes the same filename with a `.part` suffix. Both live under `~/.loxa/models`.

Before transferring a model, Loxa requests the remote size and compares it with the
size pinned in the registry. It also checks available disk space for the bytes still
needed.

When a non-empty `.part` file is smaller than the expected file, Loxa requests the
remaining range. A partial response must begin at the requested byte and report the
expected total size. If the server ignores that range and returns a complete response,
Loxa removes the partial file and restarts from byte zero.

After the response completes, Loxa checks the completed byte count and verifies the
SHA-256 digest. Only then does it rename the `.part` file to the final filename. An
already completed file is also checked for its expected size and SHA-256 digest before
it is accepted as present.

`loxa rm <id>` removes both the final file and its `.part` file when they exist.

## Hugging Face token variables [#hugging-face-token-variables]

Loxa resolves Hugging Face credentials in this order:

1. A non-empty `HF_HUB_DISABLE_IMPLICIT_TOKEN` disables token resolution.
2. A non-empty `HF_TOKEN` value is used.
3. `HF_TOKEN_PATH` names a token file when `HF_TOKEN` is not available.
4. `HF_HOME` changes the directory containing the default `token` file.
5. When `HF_HOME` is unset, `XDG_CACHE_HOME` changes the cache root, with the token at
   `$XDG_CACHE_HOME/huggingface/token`.
6. Otherwise, the default token path is `~/.cache/huggingface/token`.

Do not place a real token in documentation or command examples.

See [CLI reference](/docs/cli) for command syntax and
[Troubleshooting](/docs/troubleshooting) for download errors.