---
name: abliterated-cloud
description: Use ABLITERATED.cloud safely through its token-protected OpenAI-compatible Chat Completions API.
license: Apache-2.0
metadata:
  homepage: https://abliterated.cloud/
  repository: https://github.com/eminogrande/mn-uncensored
---

# ABLITERATED.cloud

Use this skill when a user wants to call an ABLITERATED.cloud model through an OpenAI-compatible client.

## Safety boundary

- Never invent or expose an API token.
- Obtain the token only from the user's existing secret store or environment.
- Never write a token into source code, logs, configuration committed to Git, or a public issue.
- Do not assume an unavailable model will wake automatically.
- Do not start, arm, wake, or launch `cebeuq/Ornith-1.0-397B-abliterated-W4A16` without explicit operator cost acknowledgement.
- Do not claim that "abliterated" guarantees zero refusals or correctness.

## Connection

Use the operator-provided base URL ending in `/v1`.

Authenticate with:

```http
Authorization: Bearer sk-mn-...
```

## Models

- `huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated` — Qwen3.6 35B A3B — Abliterated.
- `YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated` — Ornith 1.0 35B — Abliterated.
- `huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated` — Qwythos 9B Claude Mythos 5 — Abliterated.
- `cebeuq/Ornith-1.0-397B-abliterated-W4A16` — Ornith 1.0 397B W4A16 — Abliterated.

The exact Hugging Face repository ID is also the primary API model ID. Compatibility aliases for existing clients are `mn/god`, `mn/code`, `mn/fast`, `mn/ornith-397b`, and `nuri/ornith-397b-abliterated`.

## Supported contract

Prefer:

- `GET /v1/models`
- `POST /v1/chat/completions`
- streaming
- OpenAI-style `tools` and `tool_choice`

Do not assume full parity with every OpenAI API endpoint.

## Before a call

1. Confirm the desired public model ID.
2. Confirm the route is armed or started.
3. Confirm the user understands cold-start latency.
4. For 397B, confirm the explicit two-H200 cost acknowledgement.
5. Use a conservative `max_tokens` value for the first request.

## After testing

Tell the operator to hard-stop the model instead of relying only on the idle window.