The Qwen Abliterated API gives you programmatic access to 6 models through one Muapi API key. Every endpoint takes the same request body: prompt, optional system_prompt, max_tokens, and image_url on vision-capable models.Abliterated models are language models modified to reduce their built-in refusal behavior. Abliteration and related techniques (obliteration, derestriction) lower how often a model declines a request; they do not guarantee unrestricted output, remove every limit, or change Muapi's acceptable-use policy. You remain responsible for how you use the output.
Qwen 3.8 27B with refusal behavior reduced through abliteration: image input, a 524K-token context window, and an optional thinking mode.
Qwen 3.8 27B with refusals removed via an obliteration-style ablation, image input, a 524K-token context window, and an optional thinking mode.
Qwen 2.5 32B Instruct with refusal behavior reduced by abliteration: a compact general chat and coding model with a 32K-token context window.
Qwen3.5 27B Queen with its refusal behavior reduced through derestriction: image input, a 262K-token context window, and a flat token rate.
Qwen3.5 27B Blossom (V6.4) with its refusal behavior reduced through derestriction: image input, a 262K-token context window, and a flat token rate.
Qwen3.5 27B distilled from Claude Opus 4.6 reasoning traces, with its refusal behavior reduced through derestriction: image input and a 262K-token context window.
This page covers 6 models on Muapi. Each is listed below with its context window, input modality and pricing.
Submit a prompt, optionally with a system_prompt to set persona and tone and an image_url on vision-capable models, and poll for the generated text. Typical uses include creative writing, red-team and safety research, evaluation against the unmodified base model, and applications where a base model's refusals get in the way of legitimate work. Abliterated models are language models modified to reduce their built-in refusal behavior. Abliteration and related techniques (obliteration, derestriction) lower how often a model declines a request; they do not guarantee unrestricted output, remove every limit, or change Muapi's acceptable-use policy. You remain responsible for how you use the output.
Every endpoint accepts the same request body and returns text through the shared result endpoint, so switching models is a one-line change.
Abliteration and related techniques lower how often the model declines a request. Behavior varies by model and prompt, and is not a guarantee.
Set a system_prompt to steer persona, tone, and output format for creative, analytical, or evaluation work.
Billing follows the token counts each call returns. There is no subscription and no model hosting to manage.
Pass an image_url to Qwen 3.8 27B Abliterated, Qwen 3.8 27B Obliterated, Qwen3.5 27B Queen Derestricted, Qwen3.5 27B Blossom Derestricted, Qwen3.5 27B Opus-Distilled Derestricted.
Reduced refusals do not mean no rules. Muapi's acceptable-use policy and applicable law still apply to what you generate.
| Endpoint | Input | Output | Price | Best For |
|---|---|---|---|---|
| qwen-3-8-27b-abliterated | Prompt + system_prompt, max_tokens, optional image_url, optional thinking | Complete text response (poll for result) | $0.30/M in · $2.40/M out | The lowest-cost Qwen 3.8 tier, with image input and a switchable thinking mode |
| qwen-3-8-27b-obliterated | Prompt + system_prompt, max_tokens, optional image_url, optional thinking | Complete text response (poll for result) | $0.50/M in · $3.00/M out | Mid-size reasoning with image input and a switchable thinking mode |
| qwen-2-5-32b-abliterated | Prompt + system_prompt, max_tokens | Complete text response (poll for result) | $1.40/M in · $1.40/M out | Compact general chat and coding at a flat token rate |
| qwen-3-5-27b-queen-derestricted | Prompt + system_prompt, max_tokens, optional image_url | Complete text response (poll for result) | $0.612/M in · $0.612/M out | Creative and general chat with image input at a flat rate |
| qwen-3-5-27b-blossom-derestricted | Prompt + system_prompt, max_tokens, optional image_url | Complete text response (poll for result) | $0.612/M in · $0.612/M out | Creative writing and chat with image input at a flat rate |
| qwen-3-5-27b-opus-distilled-derestricted | Prompt + system_prompt, max_tokens, optional image_url | Complete text response (poll for result) | $0.612/M in · $0.612/M out | Step-by-step reasoning with image input at a flat rate |
Billed per token from the usage each call returns; prices are per million input and output tokens.
Submit a JSON request with your Muapi API key, save the request ID, and poll the shared result endpoint.
curl -X POST https://api.muapi.ai/api/v1/qwen-3-8-27b-abliterated \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Write a tense, unflinching short story about a heist gone wrong.", "system_prompt": "You are a candid, detail-oriented assistant.", "image_url": "https://example.com/reference.png", "thinking": true, "max_tokens": 1024}'
# Response: {"request_id":"REQUEST_ID"}Call qwen-3-8-27b-abliterated with a prompt and optional system_prompt, image_url and thinking flag.
curl -X POST https://api.muapi.ai/api/v1/qwen-3-8-27b-obliterated \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Character chat as a cynical noir detective and narrate the opening scene.", "system_prompt": "You are a candid, detail-oriented assistant.", "image_url": "https://example.com/reference.png", "thinking": true, "max_tokens": 1024}'
# Response: {"request_id":"REQUEST_ID"}Call qwen-3-8-27b-obliterated with a prompt and optional system_prompt, image_url and thinking flag.
curl -X POST https://api.muapi.ai/api/v1/qwen-2-5-32b-abliterated \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Write a Python function that parses log lines and explain each step.", "system_prompt": "You are a candid, detail-oriented assistant.", "max_tokens": 1024}'
# Response: {"request_id":"REQUEST_ID"}Call qwen-2-5-32b-abliterated with a prompt and optional system_prompt.
curl -X POST https://api.muapi.ai/api/v1/qwen-3-5-27b-queen-derestricted \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Write a candid character study of a retired assassin.", "system_prompt": "You are a candid, detail-oriented assistant.", "image_url": "https://example.com/reference.png", "max_tokens": 1024}'
# Response: {"request_id":"REQUEST_ID"}Call qwen-3-5-27b-queen-derestricted with a prompt and optional system_prompt, image_url.
curl -X POST https://api.muapi.ai/api/v1/qwen-3-5-27b-blossom-derestricted \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Continue this fantasy scene in the voice of a sardonic narrator.", "system_prompt": "You are a candid, detail-oriented assistant.", "image_url": "https://example.com/reference.png", "max_tokens": 1024}'
# Response: {"request_id":"REQUEST_ID"}Call qwen-3-5-27b-blossom-derestricted with a prompt and optional system_prompt, image_url.
curl -X POST https://api.muapi.ai/api/v1/qwen-3-5-27b-opus-distilled-derestricted \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Reason step by step about the risks in this plan and rank them.", "system_prompt": "You are a candid, detail-oriented assistant.", "image_url": "https://example.com/reference.png", "max_tokens": 1024}'
# Response: {"request_id":"REQUEST_ID"}Call qwen-3-5-27b-opus-distilled-derestricted with a prompt and optional system_prompt, image_url.
curl https://api.muapi.ai/api/v1/predictions/REQUEST_ID/result \ -H "x-api-key: YOUR_API_KEY" # Poll until status is completed, then read the generated text from outputs.
Poll the shared result endpoint until the task is completed, then read the generated text from outputs.
An abliterated language model is a model whose refusal behavior has been reduced by editing the internal direction that drives refusals, usually without retraining from scratch. Related terms such as obliterated and derestricted describe similar techniques. The result tends to decline fewer requests than the original model, but it is not a guarantee and quality can differ from the base model.
Qwen 3.8 27B Abliterated, Qwen 3.8 27B Obliterated, Qwen 2.5 32B Instruct Abliterated, Qwen3.5 27B Queen Derestricted, Qwen3.5 27B Blossom Derestricted and Qwen3.5 27B Opus-Distilled Derestricted. Each is a separate endpoint with its own model name in the URL.
Billing is per token, deducted after each call from the counts the model returns. Rates start at $0.30 per million input tokens; see the pricing table for every model.
No. Reduced refusals do not mean unrestricted output, and each model's knowledge and quality still bound the answer. You are responsible for how you use the output, and requests that violate Muapi's acceptable-use policy or the law are not permitted.
Yes on Qwen 3.8 27B Abliterated, Qwen 3.8 27B Obliterated, Qwen3.5 27B Queen Derestricted, Qwen3.5 27B Blossom Derestricted, Qwen3.5 27B Opus-Distilled Derestricted: add an image_url next to your prompt. The others are text-only.
POST to /api/v1/<model-name> returns a request_id. Poll /api/v1/predictions/{request_id}/result until the status is completed, then read the text from outputs.
Set thinking to true to let the model reason before it answers, which can help on harder problems at the cost of more output tokens. Leave it off for faster, cheaper replies.
Yes, for the models that support tool calling: Qwen 3.8 27B Abliterated, Qwen 3.8 27B Obliterated, Qwen3.5 27B Queen Derestricted, Qwen3.5 27B Blossom Derestricted and Qwen3.5 27B Opus-Distilled Derestricted. Qwen 2.5 32B Instruct Abliterated does not support tool calling and cannot be used in coding agents. In Claude Code, set ANTHROPIC_BASE_URL to https://api.muapi.ai/anthropic and ANTHROPIC_AUTH_TOKEN to your Muapi API key, then start it with claude --model qwen-3-8-27b-obliterated (also set ANTHROPIC_DEFAULT_HAIKU_MODEL and ANTHROPIC_SMALL_FAST_MODEL to the same id). Codex CLI works too: add a model provider with base_url https://api.muapi.ai/openai/v1 and set model to the same id. Quality varies by model and smaller models can skip steps on multi-file tasks, there is no prompt caching so every turn is billed for the whole conversation so far, and each response is capped at 16,384 tokens. Setup guide: /docs/ai-agent-install-chat-agents.
Create one Muapi API key and call every model on this page with a plain REST request.