ABLITERATEDAI / DEVELOPER FIELD GUIDE
Change the provider. Keep the client. Understand the differences.
An OpenAI-compatible API can reduce migration effort because your application keeps the familiar chat-completions request shape. It does not make models, limits, features, costs, or error handling interchangeable. This guide uses AbliteratedAI's currently documented text API as a concrete example.
01 / Confirm the endpoint and the model actually available
The service currently documents one text model with the public ID abliterated-235b-a22b-fp8 and base URL https://api.abliteratedai.dev/v1. The separately advertised 2.8T model is incoming, not available at that endpoint. Check service status and model documentation before integrating; do not infer capacity from a name.
02 / Python: three connection settings and a bounded request
Install openai. Create an approved account and API key, store the key in a server-side environment variable, and verify your prepaid balance before a paid request. This code is illustrative and will issue a billable request when you execute it with real credentials.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.abliteratedai.dev/v1",
api_key=os.environ["ABLITERATED_API_KEY"],
max_retries=0,
)
response = client.chat.completions.create(
model="abliterated-235b-a22b-fp8",
messages=[{"role": "user", "content": "Say hello."}],
max_tokens=128,
stream=False,
)
print(response.choices[0].message.content)
Do not use the illustrative abl-sk-... string as a credential. Never place a real key in public client-side code or a screenshot.
03 / Node: the same wire format
The Node OpenAI SDK accepts the same custom endpoint. This example also disables automatic retries; after an unknown timeout you must inspect account usage before trying again to avoid inadvertently paying twice.
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.abliteratedai.dev/v1",
apiKey: process.env.ABLITERATED_API_KEY,
maxRetries: 0,
});
const response = await client.chat.completions.create({
model: "abliterated-235b-a22b-fp8",
messages: [{ role: "user", content: "Say hello." }],
max_tokens: 128,
stream: false,
});
console.log(response.choices[0]?.message?.content);04 / Understand the GPU warm-up path
This backend can sleep when idle. The first request may return HTTP 503 with provider_warming and a Retry-After header. That particular response is generated before credit reservation. Wait the indicated interval and retry with a bounded total waiting period. Do not automatically retry a request that might already have been admitted, such as an unknown network failure or a 409 conflict.
For a tested reference with bounded waiting, open the official quickstart and download first_request.py. Its default mode runs without network calls; --send explicitly opts into a real request and possible charges.
05 / Know what compatibility does not include
The current endpoint accepts non-streaming text chat completions, with stream: false and a bounded output-token limit. Streaming, function/tool calling, web search, images, and JSON mode are not enabled. Other providers' SDK features may silently not carry over; read the supported parameters rather than assuming parity.
Evaluate application-specific answer quality, refusal behavior, prompt formats, throughput, error responses and token usage before production migration. No benchmark comparison is implied by this guide.
06 / Budget before traffic
AbliteratedAI currently offers prepaid credits and timed inference sessions. Session fees, token overage, cold starts and API usage are separate concepts. Read the live pricing panel and billing guide for current amounts; do not assume a fixed price or a continuously warm GPU. Keep your development account's limits low during the first integration.
Test the integration before migrating production.
Start with a disposable test project, a limited API key and a small token cap. Check your actual bill and the request history after the first successful response.
Create an API account → Read the quickstart → Check status →