LLM Inference API

Run open models on Nosana GPUs with the API key you already have. The API is OpenAI-compatible, so an existing client works once you change the base URL and the key.

Base URL	https://inference.nosana.com/v1
Authentication	Authorization: Bearer nos_... — the same key you use for deployments
Billing	Per token, against your credits

List models

A model appears here only while a GPU is actually serving it, so this is the list you can call right now.

curl https://inference.nosana.com/v1/models \
  -H "Authorization: Bearer nos_xxx_your_api_key"

Response:

{
  "object": "list",
  "data": [
    {
      "id": "<model-id>",
      "object": "model",
      "created": 1789399587,
      "owned_by": "nosana",
      "name": "…",
      "description": "…",
      "context_length": 16384,
      "max_output_tokens": 4096,
      "pricing": { "prompt": "0.000000100", "completion": "0.000000300" },
      "available": true
    }
  ]
}

pricing is in dollars per token. Multiply by a million for the usual per-million figure: the values above are $0.10 per million prompt tokens and $0.30 per million completion tokens.

Which models are served changes over time, so read the id from this list rather than hardcoding one. The examples below use <model-id> for whichever you pick.

Chat completions

from openai import OpenAI

client = OpenAI(
    base_url="https://inference.nosana.com/v1",
    api_key="nos_xxx_your_api_key",
)

MODEL = next(m.id for m in client.models.list().data)

reply = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "Say OK"}],
)

print(reply.choices[0].message.content)
print(reply.usage)

/v1/completions works the same way for the older text-completion shape.

Streaming

Set stream: true for server-sent events. Ask for stream_options.include_usage and the final frame carries the token counts, which is what you are billed on.

curl https://inference.nosana.com/v1/chat/completions \
  -H "Authorization: Bearer nos_xxx_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<model-id>",
    "messages": [{ "role": "user", "content": "Count to three" }],
    "stream": true,
    "stream_options": { "include_usage": true }
  }'

Reasoning models

Some models think before answering. The thinking is returned in message.reasoning, apart from the answer in message.content, so you can show it, log it, or drop it.

Those tokens are generated and therefore billed, which is worth knowing if a short answer costs more than its length suggests.

Usage and cost

Every request is recorded. Ask for your own history and lifetime totals:

curl "https://api.nosana.com/credits/llm/usage?limit=20" \
  -H "Authorization: Bearer nos_xxx_your_api_key"

Response:

{
  "usage": [
    {
      "requestId": "d757881a-c690-4c3f-a6bf-0c61835ed867",
      "model": "<model-id>",
      "promptTokens": 17,
      "completionTokens": 4096,
      "costNano": 1230500,
      "unmetered": false,
      "estimated": false,
      "streamed": true,
      "status": 200,
      "createdAt": "2026-09-15T07:54:19.585Z"
    }
  ],
  "total": 6,
  "totals": { "promptTokens": 102, "completionTokens": 9828, "costNano": 2958600 }
}

Costs are in nano-dollars: 1,000,000 nano-dollars is one credit. A request costing less than a whole credit is carried forward rather than rounded up, so small requests are not overcharged.

Errors

Errors use the OpenAI shape, so a client that already handles them needs no changes.

Status	code	Meaning
401	invalid_api_key	The key is missing, malformed, or belongs to another environment
402	insufficient_credits	Not enough credits to cover the request
404	model_not_found	That model is not currently being served
404	not_found	No such route — only the endpoints above are proxied
502	upstream_unreachable	The gateway could not reach the model
504	upstream_timeout	The model did not answer in time

{
  "error": {
    "message": "The model 'nope/nope' does not exist.",
    "type": "invalid_request_error",
    "param": null,
    "code": "model_not_found"
  }
}

A model_not_found usually means no GPU is serving that model at the moment rather than that it was removed. Call /v1/models to see what is available.

The header must be Authorization: Bearer nos_.... A bare token without the Bearer prefix is rejected as invalid_api_key.