# Unsloth

Source: https://docs.aicentos.com/en/clients/unsloth

Language: en

> Unsloth is a fine-tuning and local inference framework, not a chat client — why it cannot connect to aicentos

Unsloth is an open-source framework for model fine-tuning and local inference, with a desktop app and Unsloth Studio (a browser interface). Its `Connections` panel looks like model provider settings, but **it only accepts local or self-hosted endpoints, and cloud providers are a fixed list** — there is no any OpenAI-compatible URL entry, so inference requests cannot be pointed at aicentos.

## Why it cannot connect directly

* In Studio's `Settings → Connections → Add Provider`, the choices are **local or self-hosted services** such as llama.cpp, vLLM, and Ollama, and the Base URL you fill in looks like `http://localhost:8000/v1`.
* The cloud side is a fixed list (OpenAI, Anthropic, OpenRouter, ChatGPT/Codex) plus an API Key, with no entry for a custom endpoint.
* In the official docs, remote OpenAI-compatible endpoint means an inference service **you deploy yourself**, not a third-party aggregation gateway.
* Its direction is also server-side: Unsloth exposes an OpenAI-compatible API of its own for others to connect to.

## Configuration to fill in

| Field    | Value                                              |
| -------- | -------------------------------------------------- |
| API Key  | The key created in the console                     |
| Base URL | `https://aicentos.com/v1`                          |
| Model    | A model name from the Model Gallery in the console |

The table above lists the generic aicentos-side parameters, for use with the alternatives below.

## Alternatives

### Alternative 1: Training is training, chatting is chatting

Use Unsloth for fine-tuning and local inference, and a client such as [Cherry Studio](https://docs.aicentos.com/en/clients/cherry-studio) or [Chatbox](https://docs.aicentos.com/en/clients/chatbox) to connect aicentos for everyday conversation. The two do not interfere with each other.

### Alternative 2: Serve locally, point the front end at it

What Unsloth exposes is an OpenAI-compatible endpoint. Once the `unsloth` service is running locally, set the client's Base URL to its address (for example `http://localhost:8000/v1`) — the same set of steps as filling in aicentos.

### Alternative 3: To compare results, switch the front end

To compare a locally fine-tuned model with a general model on aicentos, a client that supports several providers side by side is the least work, switching in the same interface.

## Test

Once the local service is up, first `curl http://localhost:8000/v1/models` to confirm the endpoint is alive, then send one prompt in the client.

## Notes

* Do not treat Studio's `Connections` as a BYOK panel; what it governs is the local runtime.
* Local inference is hardware-hungry; for long-context and high-volume tasks, prefer aicentos.
* The official docs describe no workflow for generating training data with a third-party API, so do not plan a pipeline around that idea.

## Related links

* [Unsloth Connections docs](https://unsloth.ai/docs/integrations/connections)
* [unsloth repository](https://github.com/unslothai/unsloth)