Call open models with one API key
The GPUwerk Inference API serves open-weight models from DGX Sparks, priced per token and paid from your account credit. One API key works for every model, in the OpenAI format and in the Anthropic format, so most clients need only a new base URL and key. There is no machine to rent and nothing to deploy.
You need:
- A GPUwerk account with credit. Requests are paid from the same balance as rentals.
- An API key with the
inferencescope, created in the console under AI inference. The key is shown once; keep it secret.
Models and prices
Prices are in US dollars per million tokens. The console's AI inference page always shows the current list and prices; this table is from November 20, 2026.
| Model ID | Prompt | Cached prompt | Completion | Context | Max output |
|---|---|---|---|---|---|
qwen3.6-35b-a3b (Qwen3.6 35B A3B, NVFP4) | $0.05 | $0.02 | $0.70 | 131,072 | 8,192 |
The model is quantized to NVFP4 and served with vLLM. Its reasoning is returned separately from the answer (reasoning_content in the OpenAI format), not inside it. Prompt caching is automatic: a repeated prompt prefix is charged at the cached price. Native tool calling (the tools parameter) is not supported yet. GET /v1/models lists the models available to your key.
OpenAI format
Base URL https://api.gpuwerk.com/inference/v1, your key as a bearer token. Chat completions and completions work, streamed or not.
curl https://api.gpuwerk.com/inference/v1/chat/completions \
-H "Authorization: Bearer $GPUWERK_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "qwen3.6-35b-a3b",
"messages": [{"role": "user", "content": "Summarise GDPR Article 28 in three bullets."}]}'
With the OpenAI Python SDK:
import os
from openai import OpenAI
client = OpenAI(base_url="https://api.gpuwerk.com/inference/v1",
api_key=os.environ["GPUWERK_API_KEY"])
resp = client.chat.completions.create(
model="qwen3.6-35b-a3b",
messages=[{"role": "user", "content": "Summarise GDPR Article 28 in three bullets."}],
)
print(resp.choices[0].message.content)
Anthropic format
For clients that speak the Anthropic Messages API, the base URL is https://api.gpuwerk.com/inference (without /v1) and the key goes in x-api-key.
import os
import anthropic
client = anthropic.Anthropic(base_url="https://api.gpuwerk.com/inference",
api_key=os.environ["GPUWERK_API_KEY"])
msg = client.messages.create(
model="qwen3.6-35b-a3b",
max_tokens=1024,
messages=[{"role": "user", "content": "Summarise GDPR Article 28 in three bullets."}],
)
print(msg.content[0].text)
Use it from your tools
Any tool with an "OpenAI-compatible" provider setting works with three values: base URL https://api.gpuwerk.com/inference/v1, your API key, and the model ID qwen3.6-35b-a3b.
- Open WebUI: Admin settings, Connections, add an OpenAI API connection with the base URL and key. See Set up Open WebUI.
- LiteLLM: a model entry with
model: openai/qwen3.6-35b-a3b,api_baseandapi_key. See Set up LiteLLM. - SillyTavern, LibreChat, Jan, n8n and similar: choose the OpenAI-compatible or custom endpoint option and enter the same three values.
- Coding agents that need native tool calls will not work until tool calling is supported.
Limits and errors
Each account has limits, shown in the console: today 60 requests a minute, 2 requests in flight and $25 of spending per UTC day. Errors you may see:
- 402: your credit does not cover the most this request could cost. Top up in the console.
- 429: a limit was reached, or no machine is free right now. The response carries
Retry-After; retry after it. The request was not queued or charged. - 401 / 403: the key is unknown or revoked, lacks the
inferencescope, or is bound to one instance.
The API runs on Sparks while no rental needs them, and rentals always take priority, so capacity varies. There is no service level agreement. Batch jobs that can retry later suit it best. The terms are in section 21 of the Terms.
Where requests run and what is kept
A request runs on a DGX Spark that is free at that moment: one GPUwerk owns in the Czech Republic, or one owned by a hosting partner, which may be in another country. The console's AI inference page shows the countries of the machines serving requests now.
To keep a key's requests on GPUwerk-owned machines in the Czech Republic, tick "GPUwerk-owned machines in Czechia only" when you create the key in the console, or send "own_hardware_only": true when you create it with POST /v1/api-keys; PATCH /v1/api-keys/<id> changes it later (see the API reference). It is off by default. Fewer machines can take such a key's requests, so it may get 429 more often.
Prompts and completions are held in memory while a request runs. They are not stored, logged or used for training. Billing records hold the time, model and token counts, never content. The Inference API data policy describes where requests run, what protects them on a partner's machine and which sub-processors are involved.