API reference

The API follows the OpenAI chat completions format. If your code or tool already talks to an OpenAI-compatible endpoint, change the base URL to https://ablitllm.com/v1, use a AblitLLM key and pick a model id.

Quickstart

Create an account, create a key under API keys, and send a request.

curl https://ablitllm.com/v1/chat/completions \
  -H "Authorization: Bearer ab_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-27b-abliterated",
    "stream": true,
    "messages": [{"role": "user", "content": "Write a port scanner in Go."}]
  }'

Authentication

Send your key in the Authorization header as a bearer token. Keys start with ab_live_. Keep them on a server: a key in browser code can be read and spent by anyone who opens the page.

Authorization: Bearer ab_live_...

Chat completions

POST https://ablitllm.com/v1/chat/completions

Fields not listed here are ignored.

Field	Type	Meaning
model	string, required	A model id from the models table.
messages	array, required	The conversation: objects with role (system, user, assistant, tool) and content.
stream	boolean	Send the answer as server-sent events while it is generated. Default false.
stream_options	object	Set include_usage to true to receive a final chunk with token counts.
max_tokens	integer	Upper bound on generated tokens. Capped at the model's maximum and at what your balance can pay for.
temperature, top_p, top_k, min_p	number	Sampling controls, passed to the model unchanged.
stop	string or array	Sequences that end the answer.
seed	integer	Makes sampling repeatable where the model supports it.
tools, tool_choice	array, string or object	Function definitions the model may call.
response_format	object	Ask for JSON output, optionally with a schema.
presence_penalty, frequency_penalty, repetition_penalty	number	Repetition controls.

The response is a standard chat completion object, with token counts in usage.

Streaming

With stream: true the answer arrives as server-sent events, one chat.completion.chunk per event, ending with data: [DONE]. Closing the connection stops generation, and you are billed for what was produced up to that point.

Tool calls

Pass function definitions in tools. When the model decides to call one, the answer carries tool_calls instead of text. Run the function, append a message with role tool and the result, and call the API again. This is what coding agents do on every step.

Models

GET https://ablitllm.com/v1/models

Id	Context	Max output	Input / 1M	Output / 1M
qwen3.8-27b-abliterated	196,000	16,384	$0.30	$1.20
coder-abliterated (planned)	128,000	16,384	n/a	n/a

Errors

Errors use the OpenAI shape, so existing clients surface them correctly.

{
  "error": {
    "message": "Your credit balance is empty. Add credits to continue.",
    "type": "insufficient_credits",
    "code": "insufficient_credits"
  }
}

Status	Type	Meaning
401	invalid_api_key	The key is missing, wrong or revoked.
402	insufficient_credits	The balance is empty. Add credits to continue.
400	invalid_request_error	The body is not valid, or the model rejected it (for example, the prompt exceeds the context).
400	content_policy_violation	The request asks for sexual content involving minors, the one blocked category.
404	model_not_found	No model with that id.
404	model_not_available	The model is listed as planned and cannot be called yet.
429	rate_limit_exceeded	Too many requests per minute. The Retry-After header says how long to wait.
429	too_many_concurrent_requests	Too many requests in progress for this account, or the balance is already committed to them.
429	daily_limit_reached	The fair-use limit of included chat was reached (chat only).
502	upstream_error	The inference server failed. Failed requests are not billed.

Limits and billing

60 requests per minute per key.
4 requests in progress at the same time per account.
Request bodies up to 2 MB.
Each request costs prompt tokens times the input price plus generated tokens times the output price, taken from your prepaid balance when the request ends.
If the inference server does not report token counts, they are estimated from the text length and the request is marked as estimated in your usage table.

What is stored

For each request: the account, the key, the model, token counts, cost, status, duration and time. The text of prompts and answers is not written to disk or to logs by the gateway. A blocked request leaves the rule that blocked it and the time, not its text. Details in the privacy policy.