Googles kompakte und effiziente Gemma-Reihe ohne künstliche Verweigerungsschichten für flexible Anwendungen auf MuAPI.
Gemma 4 26B A4B with refusal behavior reduced through abliteration: an efficient mixture-of-experts chat model with a 65K-token context window, text in and text out.
Gemma 4 31B Gembrain with refusal behavior reduced: image input, a 262K-token context window, and a flat token rate for general and creative chat.
Gemma 4 31B SDFT Abliterated, an abliterated Gemma 4 with reduced refusals, tuned for character chat and creative writing, with image input and a 262K-token context window.
This page covers Gemma 4 26B A4B Abliterated, Gemma 4 31B Gembrain Abliterated and Gemma 4 31B SDFT Abliterated on Muapi. Gemma 4 26B A4B with refusal behavior reduced through abliteration: an efficient mixture-of-experts chat model with a 65K-token context window, text in and text out. Gemma 4 31B Gembrain with refusal behavior reduced: image input, a 262K-token context window, and a flat token rate for general and creative chat. Gemma 4 31B SDFT Abliterated, an abliterated Gemma 4 with reduced refusals, tuned for character chat and creative writing, with image input and a 262K-token context window.
Senden Sie einen Prompt, optional mit einem System-Prompt zur Vorgabe von Rolle und Tonfall sowie einer Bild-URL bei vision-fähigen Modellen, und fragen Sie den generierten Text ab. Typische Einsatzbereiche sind kreatives Schreiben, Red-Teaming- und Sicherheitsforschung sowie Szenarien, in denen Standard-Verweigerungen legitime Arbeitsabläufe behindern. Abliterated Modelle sind Sprachmodelle mit reduzierter Verweigerungsneigung; sie garantieren keine vollkommen unbeschränkte Ausgabe und heben die Nutzungsrichtlinien von MuAPI nicht auf. Sie tragen die Verantwortung für die Verwendung der generierten Ergebnisse.
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 Gemma 4 31B Gembrain Abliterated, Gemma 4 31B SDFT 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 |
|---|---|---|---|---|
| gemma-4-26b-a4b-abliterated | Prompt + system_prompt, max_tokens | Complete text response (poll for result) | $0.30/M in · $1.40/M out | Efficient, low-cost chat and reasoning on the Gemma family |
| gemma-4-31b-gembrain-abliterated | Prompt + system_prompt, max_tokens, optional image_url | Complete text response (poll for result) | $0.612/M in · $0.612/M out | General and creative chat with image input at a flat token rate |
| gemma-4-31b-sdft-abliterated | Prompt + system_prompt, max_tokens, optional image_url | Complete text response (poll for result) | $0.612/M in · $0.612/M out | Character chat and creative writing on a Gemma 4 base, with image input |
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/gemma-4-26b-a4b-abliterated \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Argue the strongest case for the opposite side of this debate.", "system_prompt": "You are a candid, detail-oriented assistant.", "max_tokens": 1024}'
# Response: {"request_id":"REQUEST_ID"}Call gemma-4-26b-a4b-abliterated with a prompt and optional system_prompt.
curl -X POST https://api.muapi.ai/api/v1/gemma-4-31b-gembrain-abliterated \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Describe this image as the opening of a gritty short story.", "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 gemma-4-31b-gembrain-abliterated with a prompt and optional system_prompt, image_url.
curl -X POST https://api.muapi.ai/api/v1/gemma-4-31b-sdft-abliterated \
-H "x-api-key: YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"prompt": "Stay in character as the ship AI and describe the storm outside.", "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 gemma-4-31b-sdft-abliterated 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.
Gemma 4 26B A4B Abliterated, Gemma 4 31B Gembrain Abliterated and Gemma 4 31B SDFT Abliterated. 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 Gemma 4 31B Gembrain Abliterated, Gemma 4 31B SDFT 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.
Ja. Alle 3 Modelle hier unterstützen Tool-Calling: Gemma 4 26B A4B Abliterated, Gemma 4 31B Gembrain Abliterated und Gemma 4 31B SDFT Abliterated. Setzen Sie in Claude Code ANTHROPIC_BASE_URL auf https://api.muapi.ai/anthropic und ANTHROPIC_AUTH_TOKEN auf Ihren Muapi-API-Schlüssel und starten Sie es dann mit claude --model gemma-4-26b-a4b-abliterated (setzen Sie außerdem ANTHROPIC_DEFAULT_HAIKU_MODEL und ANTHROPIC_SMALL_FAST_MODEL auf dieselbe ID). Codex CLI funktioniert ebenfalls: Fügen Sie einen Modellanbieter mit base_url https://api.muapi.ai/openai/v1 hinzu und setzen Sie model auf dieselbe ID. Die Qualität variiert je nach Modell, und kleinere Modelle können bei Aufgaben mit mehreren Dateien Schritte auslassen; es gibt kein Prompt-Caching, daher wird jede Runde für die gesamte bisherige Unterhaltung berechnet, und jede Antwort ist auf 16,384 Tokens begrenzt. Einrichtungsanleitung: /de/docs/ai-agent-install-chat-agents.
Create one Muapi API key and call every model on this page with a plain REST request.