Your tools. Our machines.

Use the browser playground or connect a chat-completions client.

From booking to first token

Reserve a model and a time.
Choose the dual pair, a single Spark, or the Mac Studio and a qualified model. Its context and concurrency limits apply to the whole session.
Create your session key.
Open My sessions after confirmation. Copy the key when it is shown; a replacement revokes the previous key.
Connect at your booked start.
Wait until My sessions shows Ready, then send your request. Use streaming for generation and keep a long client timeout.

Only the model your session currently runs is accessible; switch models from My sessions. Your API key stops accepting requests when the session ends.

Use your session key and replace model with the exact model ID from My sessions. These examples send the same prompt to your reserved model.

Python · streaming response

from openai import OpenAI

client = OpenAI(
    base_url="https://your-spark-lab/api/v1",
    api_key="YOUR_SESSION_KEY",
    timeout=1500.0,
)

stream = client.chat.completions.create(
    model="YOUR_RESERVED_MODEL",
    messages=[{"role": "user", "content": "Explain my deployment options."}],
    max_tokens=2048,
    stream=True,
)

for chunk in stream:
    if chunk.choices:
        print(chunk.choices[0].delta.content or "", end="")

Install with pip install openai, replace YOUR_SESSION_KEY, and run the script. Text prints as it arrives.

What to expect

Cold starts are normal.

We prepare scheduled models up to 20 minutes before their start. Immediate sessions prepare after payment. The paid clock waits for readiness, and the selected model stays loaded until you switch.

One request at a time.

The current million-token profiles permit one concurrent sequence. Wait for a request to finish or cancel it before starting another.

Keep long jobs bounded.

API calls do not silently join a multi-hour queue. Use background jobs for unattended prompts, with an explicit spending cap and deadline.

Session policy

Choose Start now when the full session is available, or reserve a later time. Preparation is excluded from paid time. Your clock starts at your scheduled start or when the model is ready, whichever is later; the full purchased duration is preserved. Customer idle time after readiness counts. Failed preparation initiates a full refund when the session is reconciled. Cancel a paid reservation at least 12 hours before its start for a full refund.

New requests stop at the booked end. Already accepted requests have up to two minutes to drain in the changeover buffer. There is no automatic paid extension. Owner takeover can interrupt a trial or paid reservation; access ends immediately and interrupted paid sessions receive a full refund. You are notified when this happens.

Switching models. During an active session you can switch to any other catalog model that fits your hardware, from My sessions or with POST /api/bookings/{id}/model. Your API key stays the same; /v1/models lists the model you run now, and a request naming another model is refused until you switch. The session end does not move: the new model's loading time comes out of your remaining time, and the usual load time is shown before you confirm. A model that fails to load keeps the session; pick another one.

Your data

Playground messages and replies are saved privately with your session until you clear the conversation or the session ends. Closing the playground or refreshing the page keeps that history. Usage logs contain timing, token counts, model, and session identifiers; they do not contain prompts. API and benchmark prompts are not saved as playground conversations.

Background prompts and results are encrypted on the machine and removed 72 hours after completion. Cancelling a running job charges elapsed generation time and returns the unused credit. Account access remains available after your inference key expires. Notifications contain status only. Use sample or redacted business data during the pilot.

This is a small hardware evaluation lab. Published context limits are deployment configurations, not a guarantee of task quality or a production availability SLA.