The GLM Abliterated API gives you programmatic access to GLM 5.3 Flash Abliterated, GLM 5.3 Abliterated and GLM 4.6 Derestricted v5 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.
GLM 5.3 Flash with refusal behavior reduced through abliteration: fast reasoning, a 1M-token context window, and optional image input.
The full-size GLM 5.3 with refusal behavior reduced through abliteration: a 1M-token context window, text in and text out.
GLM 4.6 (v5) with its refusal behavior reduced through derestriction: a 131K-token context window, text in and text out.
This page covers GLM 5.3 Flash Abliterated, GLM 5.3 Abliterated and GLM 4.6 Derestricted v5 on Muapi. GLM 5.3 Flash with refusal behavior reduced through abliteration: fast reasoning, a 1M-token context window, and optional image input. The full-size GLM 5.3 with refusal behavior reduced through abliteration: a 1M-token context window, text in and text out. GLM 4.6 (v5) with its refusal behavior reduced through derestriction: a 131K-token context window, text in and text out.
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 GLM 5.3 Flash Abliterated.
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 |
|---|---|---|---|---|
| glm-5-3-flash-abliterated | Prompt + system_prompt, max_tokens, optional image_url | Complete text response (poll for result) | $0.40/M in · $1.60/M out | Fast, low-cost reasoning with image input and very long context |
| glm-5-3-abliterated | Prompt + system_prompt, max_tokens | Complete text response (poll for result) | $2.50/M in · $4.50/M out | Full-size reasoning over very long context |
| glm-4-6-derestricted-v5 | Prompt + system_prompt, max_tokens | Complete text response (poll for result) | $0.80/M in · $3.00/M out | Mid-size GLM reasoning with a 131K-token context |
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/glm-5-3-flash-abliterated \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Analyze this chart and write a blunt executive summary of the risks.", "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 glm-5-3-flash-abliterated with a prompt and optional system_prompt, image_url.
curl -X POST https://api.muapi.ai/api/v1/glm-5-3-abliterated \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Summarize this 500-page report and flag every unsupported claim.", "system_prompt": "You are a candid, detail-oriented assistant.", "max_tokens": 1024}'
# Response: {"request_id":"REQUEST_ID"}Call glm-5-3-abliterated with a prompt and optional system_prompt.
curl -X POST https://api.muapi.ai/api/v1/glm-4-6-derestricted-v5 \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Draft a frank critique of this business plan.", "system_prompt": "You are a candid, detail-oriented assistant.", "max_tokens": 1024}'
# Response: {"request_id":"REQUEST_ID"}Call glm-4-6-derestricted-v5 with a prompt and optional system_prompt.
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.
GLM 5.3 Flash Abliterated, GLM 5.3 Abliterated and GLM 4.6 Derestricted v5. 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 $0.40 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 GLM 5.3 Flash Abliterated: 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.
Yes. All 3 models here support tool calling: GLM 5.3 Flash Abliterated, GLM 5.3 Abliterated and GLM 4.6 Derestricted v5. 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 glm-5-3-flash-abliterated (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.