choosing where to run it

You want the model. You do not want a second job.

Qwen3.8 27B is an open model anyone can download, and it is genuinely good. The hard part was never the weights. It is everything around them: the card, the drivers, the crash at two in the morning. Here is what each option really asks of you, and what it costs.

to first word 166 ms then 138 words a second, near reading speed

Three honest options

including the ones that are not us

Run it on your own machine

What is good: Nothing leaves your computer, and once the hardware is paid for the tokens are free.

What it costs you: You need a card with enough memory, and you own every crash, every driver update, and every hour spent tuning instead of building.

Good if you enjoy the hardware side and your work can wait when something breaks.

Rent a GPU by the hour

What is good: More memory than you own, available this afternoon.

What it costs you: The meter runs while the model loads, while you debug, and while you sleep if you forget to shut it down. You are still the operator.

Good for a burst of experiments with a hard end date.

Send requests to a gateway (this is us)

What is good: One key, one URL, and you pay for the tokens you actually use.

What it costs you: Your prompts leave your machine, and you depend on somebody else staying up.

Good if the model is a part of what you are building rather than the thing you are building.

What a real month looks like

published rates, arithmetic you can redo

Three ordinary workloads, priced
Work	Sent / received	You pay
A long chat Around fifty back-and-forth turns with a document in context.	0.4M in / 0.06M out	$0.28
A working day of coding help A few hundred questions, answers with code in them.	1.2M in / 0.25M out	$0.99
A batch of 500 summaries Two thousand words in, three hundred words out, each.	1.4M in / 0.20M out	$0.97

Rates are $0.40 for a million words you send and $2.03 for a million you get back, with repeated context billed at $0.10. There is no charge per request and no monthly fee, so an idle month costs nothing. Full billing terms.

Straight answers

including where we lose

Is this the cheapest way to run this model?: No. If you already own a capable card, your own machine is cheaper per token and always will be. We are cheaper than renting a GPU you leave running, and we cost nothing when you are idle.
What happens when many people call at once?: First-token time stays under a second at light load and climbs when the queue builds. We publish what we measured rather than a promise: 166 ms to first token, 138 tokens a second after that.
Do you read my prompts?: Paid traffic is processed in memory and not kept. Trial traffic is the exception and may be retained for model improvement, which is the trade for free credit.
Am I locked in?: No contract and no subscription. Credit does not expire, the API speaks the same shapes as the clients you already use, and switching away is a base URL change.

Try it before you decide

no card, no install

The first 100 accounts get $5 of credit, which is roughly 887 ordinary chat turns. Send one prompt in the browser and watch the cost tick up as it streams. If it is not better than what you have, you have lost a minute.

Try a prompt See the two-line setup Read the measurements

Trial prompts may be retained for model improvement; paid traffic is not. Data handling. Context window is 262K tokens. Rate limits exist for abuse control and are raised on request.