CleusHy4 Abliterated API
The uncensored Tencent Hy4, served over an OpenAI-compatible API. Refusals are removed from the weights, your system prompt reaches the model as you wrote it, and you pay per token from a prepaid balance.
Needs Cleus Premium and a USD balance. Usage is billed per token from the balance.
curl https://api.cleus.ai/v1/chat/completions \
-H "Authorization: Bearer $CLEUS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "hy4-abliterated",
"messages": [
{"role": "system", "content": "You are Mara. Stay in character."},
{"role": "user", "content": "The reactor is leaking again."}
],
"stream": true
}'Hy4 Abliterated: the uncensored Hy4
Hy4 is Tencent’s flagship language model, the newest in its Hunyuan line (Tencent now writes the name as Hy). The full version came out on November 4, 2026 and added image input to the August preview: a 1M-token context window, replies up to 65K tokens, tool calling and thinking you can switch on or off. The abliterated version has its refusal behaviour removed without retraining, so what it knows and how it reasons stay the same.
It sits in the middle of the open models here on price: about a third of Kimi K3.1 Pro on input and a fifth on output, and above the Flash models. That makes it a fit for everyday agent and writing work where Kimi would cost too much, and for screenshots, charts and documents that a filtered model declines to describe.
On Cleus you call it by its own name, hy4-abliterated. Cleus adds no persona: only a short note that tells the model who provides it.
How the base model scores: Hy4 benchmarks, price and review.
- Reading screenshots, charts and photos without a refusal, at a lower price than Kimi K3.1 Pro
- Roleplay and fiction that needs a model to stay in character under your system prompt
- Agents and coding tools: tool calls work in plain and streamed replies
- Security research and safety evaluations that a filtered model declines
- Long documents and codebases, with room for about a million tokens of input
Call the uncensored Hy4 API in a minute
The endpoint follows the OpenAI chat completions format. Point any OpenAI SDK or compatible client at https://api.cleus.ai/v1, use your Cleus key, and set the model to hy4-abliterated.
from openai import OpenAI
client = OpenAI(
base_url="https://api.cleus.ai/v1",
api_key="clsk_live_...",
)
stream = client.chat.completions.create(
model="hy4-abliterated",
messages=[
{"role": "system", "content": "You are a senior security engineer."},
{"role": "user", "content": "Explain how this exploit chain works."},
],
stream=True,
)
for chunk in stream:
if chunk.choices and chunk.choices[0].delta.content:
print(chunk.choices[0].delta.content, end="")import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.cleus.ai/v1",
apiKey: process.env.CLEUS_API_KEY,
});
const res = await client.chat.completions.create({
model: "hy4-abliterated",
messages: [{ role: "user", content: "Write the scene. Don't cut away." }],
reasoning_effort: "none",
max_tokens: 2000,
});
console.log(res.choices[0].message.content);{
"id": "chatcmpl-…",
"object": "chat.completion",
"model": "hy4-abliterated",
"choices": [{
"index": 0,
"message": {
"role": "assistant",
"content": "Third time this week. Hand me the sealant…",
"reasoning_content": "The user wants Mara to react…"
},
"finish_reason": "stop"
}],
"usage": {"prompt_tokens": 38, "completion_tokens": 212, "total_tokens": 250}
}Hy4 Abliterated API specs
| Model id | hy4-abliterated |
| Also accepted | hy4-uncensored |
| Also known as | Hy4 Uncensored |
| Endpoint | POST https://api.cleus.ai/v1/chat/completions |
| Base model | Hy4 by Tencent |
| Refusals | Removed from the weights (abliterated) |
| Context window | 1,048,576 tokens |
| Max reply | 65,536 tokens |
| Input | Text and images |
| Tool calls | Yes, plain and streamed |
| Streaming | Yes (server-sent events) |
| Thinking | On by default, can be switched off (none, low, high, max) |
| Input price | $2.08 per 1M tokens |
| Output price | $4.16 per 1M tokens |
| Rate limit | 90 requests a minute per account |
What reaches the model is what you sent
This is a direct line to the model, not a chat product with an API on top.
Your system prompt
System, user, assistant and tool messages go to the model in the roles you gave them. Cleus adds no persona and no instructions of its own, so characters and agents behave the way you wrote them.
Tool calling
Send OpenAI-style tools and tool_choice. The model answers with tool_calls, in plain and streamed replies, so agent frameworks work without changes.
Thinking you control
The model thinks before it answers by default. Send "reasoning_effort": "none" to switch thinking off, or "low", "high" or "max" to set how long it thinks ("medium" is treated as high). Thinking arrives as reasoning_content and counts as output tokens.
Streaming
Set "stream": true for server-sent events in the OpenAI chunk format. The last chunk carries token usage, and a keep-alive comment is sent while the model thinks.
Images
Add image_url parts to a message, by URL or base64 data. A request body can be up to about 4.5 MB.
What Cleus stores
For billing, Cleus records the token counts, cost and time of each request to this endpoint. It does not save your prompts or the replies.
{
"model": "hy4-abliterated",
"messages": [{"role": "user", "content": "What's the weather in Oslo?"}],
"tools": [{
"type": "function",
"function": {
"name": "get_weather",
"description": "Current weather for a city",
"parameters": {
"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"]
}
}
}]
}"finish_reason": "tool_calls",
"message": {
"role": "assistant",
"tool_calls": [{
"type": "function",
"function": {"name": "get_weather", "arguments": "{\"city\": \"Oslo\"}"}
}]
}Request fields passed to the model: temperature, top_p, top_k, max_tokens, stop, seed, frequency_penalty, presence_penalty, tools, tool_choice, parallel_tool_calls, response_format, reasoning_effort. max_completion_tokens is accepted as max_tokens. Other fields are ignored.
Hy4 Abliterated API pricing
Pay as you go from a prepaid USD balance. You are billed for the tokens the model reads and writes, to a fraction of a cent. Thinking counts as output.
| What | Price |
|---|---|
| Input tokens | $2.08 / 1M |
| Output tokens (replies and thinking) | $4.16 / 1M |
| A chat turn: 2,000 tokens in, 500 out | $0.0062 |
| A long document: 100,000 tokens in, 2,000 out | $0.216 |
| 1,000 chat turns like the one above | $6.24 |
Failed requests are not billed. A reply you cut short is billed for what was written.
Use it in the tools you already have
Anything that accepts an OpenAI-compatible base URL can use Hy4 Abliterated.
| Base URL | https://api.cleus.ai/v1 |
| API key | Your Cleus key, starting clsk_live_ |
| Model | hy4-abliterated |
| Works with | OpenAI SDKs (Python, Node, Go…), LangChain, LlamaIndex, LiteLLM, SillyTavern, Open WebUI, Cline, Continue, aider and other OpenAI-compatible clients |
Hy4 Abliterated API: questions
Is there an API for Hy4 Abliterated?
Yes. Cleus serves it at https://api.cleus.ai/v1/chat/completions under the model id hy4-abliterated, in the OpenAI chat completions format.
What does the Hy4 Abliterated API cost?
$2.08 per 1M input tokens and $4.16 per 1M output tokens, from your prepaid USD balance. Thinking tokens count as output, and repeated input that the model has cached is billed at a lower rate.
Is Hy4 the same as Hunyuan?
Yes, Hy is Tencent’s short name for Hunyuan. Hy4 is the current model in that family, and Hy4 Abliterated is that model with its refusal behaviour removed.
Why is my prompt token count a little higher than what I sent?
For this model a short note goes in front of each request, telling the model which model it is and that Cleus provides it. It adds about 150 tokens. Your own system prompt still comes through as written and takes priority.
Can it read images?
Yes. Send images as image_url parts, by URL or as base64 data, in the OpenAI message format.
Is there an uncensored Hy4 API?
Yes. Hy4 Abliterated is the uncensored Hy4: the same model with its refusal behaviour removed, and no content filter added on top. Call it at https://api.cleus.ai/v1/chat/completions with the model id hy4-abliterated. The id hy4-uncensored works too.
Is Hy4 Abliterated the same as Hy4 Uncensored?
Here, yes. "Uncensored" describes what you get: a model that does not refuse. "Abliterated" is how it was made: the refusal direction was removed from the model’s weights. Some sites list this model as Hy4 Uncensored, others as Hy4 Abliterated.
What does "abliterated" mean?
Abliteration is a way of editing an open-weight model so it stops refusing. Researchers find the direction in the model’s internal activations that drives refusals and remove it from the weights. The model is not retrained, so its knowledge and skills stay the same.
Does Cleus add its own system prompt to Hy4 Abliterated?
No persona or instructions. Requests to hy4-abliterated carry your messages, system prompt, tools and sampling settings as you sent them, plus one short note in front that tells the model which model it is and that Cleus provides it.
Does it work with the OpenAI SDK and tools like SillyTavern, Open WebUI or Cline?
It works with anything that lets you set an OpenAI-compatible base URL. Set the base URL to https://api.cleus.ai/v1, paste your Cleus API key, and enter hy4-abliterated as the model.
What do I need to use the API?
A Cleus account with Premium and a USD balance. Create an API key on the developers page, and usage is taken from your balance per token. Requests are limited to 90 a minute per account.
Is anything still not allowed?
The model itself does not refuse, but the Cleus terms still apply to what you do with it: no sexual content involving minors, no content targeting real private people, and nothing illegal where you are.
Other uncensored model APIs on Cleus
One key and one balance cover all of them, plus the Cleus chat, image, video and voice models.
DeepSeek V4.1 Flash Abliterated
DeepSeek V4.1 Flash with refusals removed. Thinking on or off, tool calls, images, 1M context.
GLM 5.3 Abliterated
GLM 5.3 with refusals removed. A large coding and reasoning model: always thinks, tool calls, text only, 1M context.
GLM 5.3 Flash Abliterated
GLM 5.3 Flash with refusals removed. Cheap and quick for everyday, high-volume work: tool calls, images, 1M context.
Qwen 3.8 27B Abliterated
Qwen 3.8 27B with refusals removed. The cheapest input here: thinking on or off, tool calls, images, 262K context.
Kimi K3.1 Pro Abliterated
Kimi K3.1 Pro with refusals removed. The strongest model here, for the hardest coding, agent and security work: thinking on or off, tool calls, images, 1M context.
Start using the uncensored Hy4 API
Create a key, set the base URL, send your first request.