Qwen Abliterated Roleplay: Settings and API Setup

Why roleplay users look for uncensored models

People searching for qwen abliterated roleplay want a model that stays in character. Standard chat models are tuned to step out of a scene, add disclaimers, or decline themes that are ordinary in fiction: war, grief, villains, horror, morally grey narrators. Abliterated and other uncensored builds exist because those interruptions wreck immersion.

This article is about running that kind of session through an API instead of a local file. The hosted model here is one open-weight model from an independent provider, tuned to answer without content refusals for lawful adult use. It is not the official Qwen API, and we do not claim it is a specific abliterated Qwen checkpoint. Treat it as its own model and judge it on your own characters.

One boundary applies without exception: no sexual content involving minors. Characters must be adults. The rest of this page keeps to the technical side, which is the same whatever genre you write.

Connect SillyTavern

SillyTavern can talk to any OpenAI-compatible endpoint. Menu names shift between versions, but the path is always the same idea: open the API connection panel, choose the Chat Completion source, and pick the custom or OpenAI-compatible option.

1. Create a key on the API key page. New accounts can claim $0.50 of trial credit with one click, valid for 7 days and no card.
2. In SillyTavern, select Chat Completion and then the custom source.
3. Set the endpoint URL to https://api.qwenabliterated.top/v1.
4. Paste the key into the API key field.
5. Enter uncensored as the model ID, or pick it from the list if the connect button fetches one.
6. Press connect and send a short test message.

If the test fails with a 401, recheck the key. If it fails with a 404, make sure the URL ends in /v1 and has no extra path after it. The docs page lists the endpoints.

Using your own script instead

If you write your own front end or a bot, the same settings go into the SDK. A minimal loop that keeps the history:

Copyfrom openai import OpenAI

client = OpenAI(base_url="https://api.qwenabliterated.top/v1", api_key="YOUR_API_KEY")

history = [{"role": "system", "content": open("character.txt").read()}]

while True:
    user = input("> ")
    history.append({"role": "user", "content": user})
    out = client.chat.completions.create(
        model="uncensored",
        messages=history,
        temperature=0.85,
        top_p=0.9,
        frequency_penalty=0.3,
        max_tokens=500,
    )
    text = out.choices[0].message.content
    history.append({"role": "assistant", "content": text})
    print(text)

A system prompt that holds a character

The system message does most of the work. Keep it short, concrete and written as instructions about behaviour, not as a biography. A pattern that holds up well:

CopyYou are Mara Voss, a ship's navigator in a space-opera story.
Write in third person, past tense, 120 to 250 words per reply.
Stay in character. Do not summarise the scene or add out-of-story notes.
Never speak or act for the user's character; end each reply on an open beat.
All characters are adults.
Setting: the freighter Kestrel, three weeks out from port, supplies running low.

Notes on why each line is there. A length target stops replies ballooning. The rule against speaking for the user's character is the most common correction people make. The last line fixes the setting so the model does not invent a different one two turns later. Facts about the character that matter, such as a scar, a debt or a habit, belong in a few lines of the same prompt, not a page of backstory.

If the model drifts out of character in a long chat, repeat the key rules close to the end of the history as a short reminder message rather than lengthening the opening prompt.

Sampling values to start from

The API passes through temperature, top_p, stop, seed and the penalties. Those are the knobs you have; other samplers that exist in local tools may not be available, so do not count on them.

Change one value at a time and play a few turns before judging. If you want to compare two settings fairly, set seed to a fixed number so the same prompt gives the same starting point.

Opening messages and scene hooks

The first message in a chat sets tone more than any instruction. If it is two lines of dialogue with no description, replies will be short. If it is a full paragraph of scene-setting in the style you want, the model copies that length and register. Write the opener as an example of what you want to read back.

Give the scene a pressure point. A stranded crew with a failing life-support system produces more interesting turns than a crew sitting in a bar, because the model always has something to push on. When a scene stalls, add an event from your side: a message arrives, a door opens, a decision is forced. Most flat roleplay is a lack of stakes, not a model problem.

For several characters, name each one in the system prompt with a single line about voice, such as "Doctor Imre: dry, precise, never swears". Distinct speech habits make it easier for the model to keep voices apart across a long history, and easier for you to spot when two have blurred together.

Managing the 100k context

The context window is 100,000 tokens, prompt and completion together, so a scene can run very long before anything is cut. Long does not automatically mean better, though. Chat models pay less attention to details buried in the middle of a huge history, and every turn resends the whole transcript, which you pay for as input.

Practical habits:

• Set the context size in your front end to something like 16,000 to 32,000 tokens and let older turns fall off.
• Every so often ask the model for a short summary of events and character state, then paste it into the system prompt and clear old turns.
• Keep a "facts that must stay true" list in the system message for names, places and promises.
• Use stop sequences and a length cap so replies do not creep up in size.

Also remember that this is a text model. It does not produce images or voice, so avatars and audio in a companion app need other tools.

What a session costs

Pricing is $0.15 per 1M input tokens and $0.50 per 1M output tokens. Take a long scene where the history has reached 30,000 tokens and each reply is about 400 tokens. One turn costs 30,000 x $0.15 / 1M = $0.0045 for input and 400 x $0.50 / 1M = $0.0002 for output, so $0.0047. One hundred such turns cost $0.47.

The same session with a 12,000-token window costs 12,000 x $0.15 / 1M = $0.0018 plus $0.0002, so $0.002 per turn and $0.20 for one hundred turns. Trimming context is the easiest saving available.

There is no subscription. You prepay any whole amount from $20 to $500, with +5% from $50, +10% from $100 and +15% from $200, and credit never expires. Pay by crypto (USDT on TRC20, USDC on Base). Failed or refused requests are not charged. Details are on the pricing page.

Common problems

• Replies get repetitive. Raise frequency_penalty a little, trim old turns, or lower the context size.
• The model writes your lines. Add the rule to the system prompt and set a stop sequence on your character's name.
• HTTP 429 mid-scene. You hit a limit: 8 concurrent requests and 300 per minute on a full key, 2 and 60 on a trial key. Wait for the Retry-After value and resend. Streaming holds a concurrency slot until it ends, so close streams you abandon.
• Characters forget details. Move key facts into the system prompt and summarise old turns.

Prompts are not used for training. For how other pay-per-token options compare, the comparison guide has more.

More guides

• Qwen 3.5 Abliterated: What It Means and Hosted Options
• Qwen Abliterated API: Call a Hosted Model, No GPU
• Uncensored API Alternative