API Documentation

Authentication

Authorization: Bearer sk-st-your-key-here

Base URL

https://api.staik.se/v1

Chat Completions

curl https://api.staik.se/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-st-your-key" \
  -d '{
    "model": "gemma4:31b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Audio transcription

Transcribe audio to text via KB-Whisper large — the Swedish National Library's Swedish-trained Whisper model (50,000 hours of Swedish speech). OpenAI-compatible — just point the openai SDK at api.staik.se.

curl https://api.staik.se/v1/audio/transcriptions \
  -H "Authorization: Bearer sk-st-your-key" \
  -F file=@moten.mp3 \
  -F model=kb-whisper-large \
  -F language=sv

Python (OpenAI SDK drop-in):

from openai import OpenAI
client = OpenAI(api_key="sk-st-your-key", base_url="https://api.staik.se/v1")

with open("moten.mp3", "rb") as f:
    result = client.audio.transcriptions.create(
        model="kb-whisper-large",  # eller "whisper-1" — båda routar till KB-Whisper large
        file=f,
        language="sv",
    )
print(result.text)

Token consumption

Audio transcription is billed at 100 tokens per second of audio from the same token pool as the chat models. That means 1 minute of audio costs 6,000 tokens and 1 hour costs 360,000 tokens.

Every response includes headers X-Audio-Duration-Seconds and X-Audio-Tokens-Charged so you can track usage per request. Max file size: 25 MB (~30–45 min of audio). Supported formats: mp3, wav, m4a, ogg, flac, webm, mp4. response_format: json (default), text, or verbose_json (with segments).

Models

Choose model via the model field in your request. Each model runs on dedicated GPU hardware in Sweden.

Chat models use Q4_K_M quantization, support streaming (SSE), and have a 262K token context window (32K for qwen3.5:9b). qwen3.6:35b-a3b and gemma4:31b have built-in vision — send images via image_url in messages. bge-m3:latest is an embedding model returning 1024-dimensional vectors (configure pgvector as vector(1024)), called via POST /v1/embeddings. kb-whisper-large is KB's Swedish-trained Whisper model, called via POST /v1/audio/transcriptions.

Plans

Error codes

Rate limit headers

Every response includes headers showing your current token usage.