Skip to main content
Pick the path that matches what you came for. You can do the other one afterwards from the same account.
Two ways to use RunInfraone account does bothModel APIsCall a model we already hostYou send an HTTP request. There is noGPU to pick and nothing to provision.Base URLhttps://api.runinfra.ai/v1Five models are hosted today.Agent optimizationHand us a model in the chatYou name a Hugging Face model and saywhat you want done with it.You end withA measured result, then an endpointor a kit you run yourself.Both draw on one workspace balance. Signing up is free and grants $1 once per account.
Two ways to use RunInfraone account does bothModel APIsCall a model we already hostYou send an HTTP request. There is noGPU to pick and nothing to provision.Base URLhttps://api.runinfra.ai/v1Five models are hosted today.Agent optimizationHand us a model in the chatYou name a Hugging Face model and saywhat you want done with it.You end withA measured result, then an endpointor a kit you run yourself.Both draw on one workspace balance. Signing up is free and grants $1 once per account.

Call a hosted model

About five minutes. You end with a working request against a model RunInfra already optimized and hosts.
1

Create an account

Sign up at runinfra.ai/sign-up with GitHub or Google. No card required, and $1 in credit lands on your first workspace for exactly this.
2

Create an API key

Go to Settings > API Keys, create a key, and store it. Keys are shown once.
3

Send your first request

The endpoint is OpenAI-compatible, so any client that speaks the OpenAI API works once you change the base URL.
max_tokens is 16,384 on purpose for this model. Reasoning tokens are billed output that count toward the budget, so a small budget can be spent entirely on thinking and return empty content you were still charged for. See Reasoning models.
4

Pick a different model

Live today: deepseek-v4-flash, deepseek-v4-pro, qwen3-8-27b, qwen3-8-2-4t-a95b, nemotron-3-5-lightning-30b, and ornith-1-5-35b. GET /v1/models returns the live list and any published context window, output cap, and limits. The Model Library shows only the prices and capabilities currently published for each model.

Streaming

Token-by-token responses over SSE.

Tool calling

A complete multi-turn tool loop.

Structured output

JSON that matches your schema.

Optimize your own model

About ten minutes of your attention. The run itself executes server-side, so you can close the tab and come back.
Quote, hold, settleruns server-side, you can close the tab01 QuoteThe plan is priced before it runsYou see the estimated charge, the hold, and your balance before you accept.02 HoldThe hold leaves your balanceIt leaves the moment the run starts, so two runs cannot spend one dollar twice.03 MeasureCandidates run on real GPUsThe baseline is measured inside the same run, on the same instrument, first.04 VerdictOne quality verdict is recordedMeasured, no quality risk, or not measured with the reason stated.05 SettleYou are charged what it provedThe rest of the hold returns. A cancelled or failed run releases all of it.Deploy, or take the kitThe estimate you accepted is a ceiling: the final charge is never morethan the number on the card you accepted.
Quote, hold, settleruns server-side, you can close the tab01 QuoteThe plan is priced before it runsYou see the estimated charge, the hold, and your balance before you accept.02 HoldThe hold leaves your balanceIt leaves the moment the run starts, so two runs cannot spend one dollar twice.03 MeasureCandidates run on real GPUsThe baseline is measured inside the same run, on the same instrument, first.04 VerdictOne quality verdict is recordedMeasured, no quality risk, or not measured with the reason stated.05 SettleYou are charged what it provedThe rest of the hold returns. A cancelled or failed run releases all of it.Deploy, or take the kitThe estimate you accepted is a ceiling: the final charge is never morethan the number on the card you accepted.
1

Fund the balance

A run is quoted before it starts and held against your balance. If the balance cannot cover the quote, nothing starts and no GPU time is spent. Add funds at Settings > Cost. See Pricing and credits.
2

Describe what you want

Open the dashboard and say it in plain words. Use @ to pick a model inline.
Continue the conversation to change anything:
Before the first run the composer asks for the two facts it cannot infer: which model, as a Hugging Face repo id, a catalog model, or weights you uploaded, and which output format, as a managed endpoint, a self-host kit, Kubernetes manifests, or a local runtime. Engine, GPU, and placement stay on Auto unless you set them.
3

Review the quote, then accept

The Plan ready card shows the estimated charge, the temporary hold, and your available balance. Accepting locks the plan and the price you agreed to, so read it before you accept. To change your mind afterwards: cancel, revise, review, accept the replacement. See Execution lifecycle.
4

Keep the winner

When the run finishes you get measured results with one recorded quality verdict, not a marketing number. Deploy the winning configuration as an endpoint, or download the kit for the output format you chose and serve it yourself. See Deployments.
Deploying a managed GPU endpoint unlocks at $50 in lifetime credit purchases. Optimizing, exporting a kit, and calling the Model APIs work on every account.

The guided prompt

The four things the agent needs before it can act.

Optimization runs

What happens once the run starts.

Deployments

Serve the winner, or take the kit.