Get a token
Request private-beta access on Signal. Every tester receives a separate, revocable Bearer token.
Private beta · four exact Hugging Face models
Intelligence, freed.Uncensored, abliterated AI.
Private API access to exact uncensored and abliterated Hugging Face models—without downloading hundreds of gigabytes or managing GPU servers yourself.
OpenAI-compatible · Revocable API tokens · Open-source infrastructure
WHAT THIS IS
Abliterated open models. One simple API.
ABLITERATED.cloud is a small, transparent hosting service for open model checkpoints with reduced refusal behavior. We run the models on cloud GPUs and expose the familiar OpenAI Chat Completions interface.
You choose the exact Hugging Face repository. We publish the model revision, hardware, context limit, route status and hourly price. No hidden model substitutions and no unprotected public endpoint.
HOW IT WORKS
Three steps. Nothing proprietary.
Choose a model
Select one of the exact Hugging Face checkpoints listed below. Internal API shortcuts are only routing identifiers.
Use your client
Connect Hermes, Pi, OpenCode, cURL or an OpenAI-compatible SDK to the same API base URL.
MODEL CATALOG
The exact Hugging Face models.
These are the complete upstream repository names—not invented ABLITERATED.cloud product names. Click any name to inspect its weights and model card directly on Hugging Face.
Qwen3.6 35B A3B — Abliterated
Hugging Face repositoryhuihui-ai/
A 36B-parameter mixture-of-experts Qwen3.6 checkpoint modified through abliteration to reduce refusal behavior. It is the general-purpose option in this catalog. The upstream model is multimodal; our current endpoint serves text.
Open this repository on Hugging Face ↗- Weights
- Free download
- License shown on HF
- Apache-2.0
- Parameters
- 36B MoE
- Hosted context
- 131,072 tokens
- Hosted on
- 1 × H200
- Managed price
- $5.45/hour
- API model ID
huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated
Ornith 1.0 35B — Abliterated
Hugging Face repositoryYuYu1015/
A 35B mixture-of-experts Ornith checkpoint for reasoning and coding, with thinking support and reduced refusals. This is the compact coding-focused choice in the current catalog.
Open this repository on Hugging Face ↗- Weights
- Free download
- License shown on HF
- Apache-2.0
- Architecture
- 35B MoE · BF16
- Hosted context
- 131,072 tokens
- Hosted on
- 1 × H200
- Managed price
- $5.45/hour
- API model ID
YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated
Qwythos 9B Claude Mythos 5 — Abliterated
Hugging Face repositoryhuihui-ai/
The smallest and least expensive model here: an abliterated Qwythos checkpoint tagged upstream for long context, function calling, tool use and agent workflows.
Open this repository on Hugging Face ↗- Weights
- Free download
- License shown on HF
- Apache-2.0
- Parameters
- 10B
- Hosted context
- 131,072 tokens
- Hosted on
- 1 × L40S
- Managed price
- $2.34/hour
- API model ID
huihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated
Ornith 1.0 397B W4A16 — Abliterated
Hugging Face repositorycebeuq/
The largest model in the catalog: a roughly 196 GB, 397B-total mixture-of-experts checkpoint for multimodal reasoning, coding and agentic work. W4A16 quantization makes deployment on two H200 GPUs practical.
Open this repository on Hugging Face ↗- Weights
- Free download
- License shown on HF
- MIT
- Parameters
- 397B total · ~17B active
- Context
- 262K upstream · 32K profile
- Hardware profile
- 2 × H200
- Managed price
- $10.90/hour
- API model ID
cebeuq/Ornith-1.0-397B-abliterated-W4A16
Free means the model weight files have no purchase price. Managed cloud inference is paid because the GPUs, storage and operation cost money. Model licenses belong to their upstream repositories and should be reviewed before commercial use. Exact pinned revisions and technical caveats are listed in our public model record.
MODEL FIELD NOTES
Four models. Four very different stories.
Trace each checkpoint from its upstream architecture through post-training, abliteration, quantization and the exact artifact in our catalog.
36B MoE · Qwenhuihui-ai/
YuYu1015/
huihui-ai/
cebeuq/
Open all model field notes
huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated
A multimodal mixture-of-experts model turned into a refusal-reduced proof of concept by huihui-ai.
Read the field note → 35B MoE · Agentic codingYuYu1015/YuYu1015-Ornith-1.0-35B-abliterated
DeepReinforce's self-scaffolding coding model, followed by YuYu1015's corrected weights-only abliteration.
Read the field note → 9B class · Long contexthuihui-ai/Huihui-Qwythos-9B-Claude-Mythos-5-1M-abliterated
The compact, inexpensive branch of the catalog, with an unusually long lineage encoded in its name.
Read the field note → 397B total · W4A16cebeuq/Ornith-1.0-397B-abliterated-W4A16
A nearly 196 GiB streaming abliteration and quantization project built for two-node inference.
Read the field note →THE API
Works with the tools you already use.
The gateway follows the OpenAI Chat Completions format. Change the base URL, add your Bearer token and send the exact API model ID. Streaming and structured tool calls are supported.
Hermes · Pi · OpenCode · OpenAI SDKs · cURL · compatible applications
Read the complete integration guidePOST /v1/chat/completions
{
"model": "YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated",
"messages": [
{ "role": "user", "content": "Build it." }
],
"stream": true
}TRANSPARENT BY DEFAULT
You can inspect what you are using.
Exact revisions
Every route is pinned to a specific public Hugging Face commit.
Protected access
Every user gets a revocable token. The model servers remain private.
Scale to zero
When an endpoint is armed, its GPU container stops after the configured idle period.
Open source
The deployment code, gateway, CLI, tests and documentation are public on GitHub.
Billing and checkout are not live yet. Published prices · access by invitation.
PLAIN ANSWERS
Frequently asked questions.
What do “abliterated”, “uncensored” and “jailbroken” mean?
Abliterated means model weights were modified to reduce learned refusal behavior. Jailbroken is a broader informal term for bypassing original restrictions. Uncensored generally means fewer built-in content refusals. None of these terms guarantees zero refusals, truth, correctness, legality or safety.
Are the models completely free?
The model weights are free to download from Hugging Face under the license shown on each repository. You can self-host them. Our managed cloud inference is paid because H200 and L40S GPU time, storage and operations cost money.
Do the models download to my computer?
No. Hugging Face stores the source files and cloud volumes cache them near the GPU servers. Your computer only runs the client or lightweight CLI.
Why does the 397B profile use 32K context?
The upstream model advertises up to 262K tokens. Our initial 397B profile uses 32K to control GPU memory and cost. This is a hosting limit, not a renamed model or a claim about the upstream checkpoint.
Does another user change the hourly price?
Not automatically. Multiple users can share one running model container. More traffic can keep it active longer, and separate model routes may run at the same time.
Is this already a public paid platform?
No. It is currently a private evaluation service. Customer accounts, quotas, metering, billing and invoicing are still roadmap work.
PRIVATE BETA
Try an exact model.
Know what is running.
Tell us which Hugging Face repository you want to use and what you want to test.
Request access on Signal