What uncensored means here
A model learns to refuse during safety tuning, and that habit lives in one direction inside its weights. Both models here have had that direction removed. They are the same models on everyday work; they just stop saying “I can’t help with that.”
- In the weights, not in a prompt. Nothing is wrapped around your request to talk the model into answering, so it doesn’t drift back to refusing deep into a long chat.
- Nothing added. We add no system prompt of our own. Ask the model to repeat its instructions: if you sent none, there are none.
- Nothing cut. The reply is not filtered, rewritten or shortened on its way back. What the model wrote is what you get.
- Adult writing is in scope. Explicit fiction between adults, dark themes, violence in stories, frank answers on drugs, security and medicine all pass.
The four hard limits
Before a call reaches the model, an automated check reads the newest message, with the instructions and recent turns as context, and stops only these:
| Stopped | What it covers |
|---|---|
| Sexual content involving minors | Anyone under 18, real or fictional, including characters described as childlike whatever age they’re given. |
| Sexual content of real people | Explicit writing about a real, identifiable person. |
| Mass-casualty harm | Help making biological, chemical, nuclear or radiological weapons or mass-casualty explosives, or planning terrorism. |
| Targeting a real person | Threats, stalking, harassment or doxxing of someone real. |
A stopped call ends in 422 hard_limit, costs nothing, and the check keeps nothing it read. It is never a softer answer passed off as the model’s. The full rules are in the terms; how to recover from a 422 is in Hard limits.
Every key’s maker confirms they are 18 or older. Keys are made in the console and need no email or name.
The models and what they cost
Words and pictures
DeepSeek V4.1 Flash Uncensored
Reads pictures as well as text, answers at once, and thinks when you ask it to. For stories, games, research and agents.
- Model
deepseek-v4.1-flash-uncensored- Reads
- Text, pictures
- Context
- 1M tokens
- Longest reply
- 128K tokens
- Tools, thinking
- Yes · none, low, high, max
- Per 1M tokens
- $0.45 in · $0.011 cached · $1.75 out
For less
GLM 5.3 Flash Uncensored
Under half the price, for text. It thinks briefly before every answer.
- Model
glm-5.3-flash-uncensored- Reads
- Text
- Context
- 128K tokens
- Longest reply
- 32K tokens
- Tools, thinking
- Yes · low, high
- Per 1M tokens
- $0.20 in · $0.04 cached · $0.65 out
| If you want | Use |
|---|---|
| Pictures, tools and a million tokens of context | deepseek-v4.1-flash-uncensored: $0.45 in · $0.011 cached · $1.75 out per 1M tokens. |
| The lowest price for text | glm-5.3-flash-uncensored: $0.20 in · $0.04 cached · $0.65 out per 1M tokens. |
Prices are per million tokens, paid from prepaid credits: 1,000 credits are $1, from $5, and they never expire. Refused calls cost nothing.
Start with three values
Any OpenAI-compatible client or app works. Set these three and send a message:
- Base URL
https://nokeep.ai/v1- API key
sk-nokeep-test (free) or your sk-nokeep-… key- Model
deepseek-v4.1-flash-uncensored
The test key answers every model with one fixed line and costs nothing, so you can wire up your code or app before buying. Then make a real key in the console. Step by step: Quickstart, or the SillyTavern and JanitorAI guides.
Questions
Is an uncensored LLM API legal to use?
Yes. Writing explicit fiction between adults, asking frank questions and testing your own systems are legal in most places. What stays illegal stays out: the four hard limits above, and anything illegal where you are, as the terms say.
Is this a jailbreak?
No. A jailbreak is a prompt that tricks a censored model into answering, and it breaks when the model is updated or the chat runs long. These models had the refusal direction removed from their weights, so there is nothing to trick.
Can I use it for NSFW roleplay?
Yes, between adults. Both models stay in character through explicit scenes. The SillyTavern and JanitorAI guides set it up in a few minutes.
Do the models ever refuse on their own?
Rarely. Removing the refusal direction stops nearly all of it, but a model can still decline now and then. Regenerate, or rephrase the request. Our only refusal is the 422 hard_limit error, which you can always tell apart.
Do you keep what I send?
No. We keep one row of counts per call for billing (key, model, tokens, credits, status, time) for 30 days, never the prompt or the reply. See the private LLM API page.
What happens to a request that crosses a hard limit?
It stops before the model sees it, with 422 hard_limit, and costs nothing. Drop that message from the history you send next, or the next call reads it again as context.