Fast inference for open models.

Open-weight LLMs on dedicated data-center GPUs, behind an OpenAI-compatible API. Your prompts are never stored.

POST /v1/chat/completions

curl https://api.koiray.com/v1/chat/completions \
  -H "Authorization: Bearer $KOIRAY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3.8-27b",
    "stream": true,
    "messages": [
      {"role": "user", "content": "Hello"}
    ]
  }'

# Or point any OpenAI SDK at this base URL

Models and pricing

Pay per token, billed in USD. Prices are per million tokens.

Qwen3.8 27B

qwen/qwen3.8-27b
FP8, official Qwen checkpoint
64K context, 16K max output

Input

$0.04/ 1M

Cached input

$0.018/ 1M

Output

$1.35/ 1M

Listing on OpenRouter is in progress. For direct API access, email info@koiray.com.

Everything agents need

The features coding tools and agent frameworks rely on, tested before every release.

Streaming first

Tokens are sent the moment they are generated. Long reasoning phases keep the connection alive, so clients never time out.

Tool calling

OpenAI-format tools and tool_choice, parsed server-side into structured calls.

Structured outputs

Constrain responses to your JSON Schema with response_format.

Reasoning, on or off

Thinking output arrives in a separate reasoning field. Turn it off per request when you only need the answer.

Your data stays yours

A short policy, because there is little to say.

No prompt or completion storage: Request and response content is processed in memory and discarded when the request ends.
No training: We never use customer data to train or fine-tune models.
Metadata only, kept 90 days: For billing and reliability we log request ID, time, model, status, token counts, cost and latency.

How we run it

Models run on vLLM on dedicated NVIDIA data-center GPUs in Tier 3+ facilities in the United States. When capacity is full we return a 429 immediately instead of queueing, so routers can fail over without waiting.