Pick a model. Click launch.

An OpenAI-compatible endpoint in your own AWS account, paid with the credits you already have, live in minutes. Then we tune your model for its chip to get 2× the tokens per dollar of stock vLLM.

Launch an endpointBring a custom model

Two ways to start

Preset

Plug and play

Choose a model we've already tuned, set how many people use it and how fast it must stream. We pick the hardware and hand you an OpenAI-compatible endpoint.

  • Live in minutes, in your AWS account
  • HTTPS at a inferoptimal.com address; change base_url, keep your code
  • Stop or terminate from this page
Configure a preset

Custom

Your model, tuned for you

Fine-tunes, new architectures, speech, embeddings. We tune your model for the chip it runs on, check outputs against your current build, and add it to your catalog as a one-click preset.

  • Same outputs, checked on your data
  • Any engine: vLLM, SGLang, llama.cpp, custom
  • Ships as a one-click preset in your catalog
Start a custom model

Launch an endpoint

AWS accountConnect to see what your account can run. No Inferoptimal account needed.
Speed per user
ModelHardware for your loadYour AWSAWS cost

Connecting adds one CloudFormation stack to your account. Its role can see GPU quotas, file quota requests, and launch, stop and terminate only the instances it tags velobyte=1. It cannot read your data or touch anything else. Each endpoint accepts traffic only from Cloudflare, which serves it over HTTPS at its own inferoptimal.com address. Delete the stack to disconnect.

Where it runs

Your AWS account

Pay with your credits

The instance runs in your account, so AWS bills it to the credits you already have. Your data stays with you.

Hardware
Billed by AWS, credits apply
Inferoptimal fee
None
Setup
One click in CloudFormation
Connect AWS

Inferoptimal cloud

Keep serving after your credits

Credits running low? Keep serving on Inferoptimal GPUs at half the AWS price. Same URL, same key, no code changes.

H100, per GPU-hour
$3.44 $6.88 on AWS
Your base URL
Stays the same
Billing
Per running hour, by card
Talk to us

Custom model

What you send

  • The model. Weights on Hugging Face, S3 or in a container, plus model code if it isn't a standard architecture.
  • How you serve it today. Engine and version, launch command, instance type.
  • Your traffic. A few hundred real requests, or typical input and output lengths, concurrency and your latency target.
  • What "same output" means. Identical tokens, or an eval set and a tolerance.

What you get

  • A preset in your catalog. Launch it from this page like any other model, on our cloud or in your AWS account.
  • Before and after on your traffic. Throughput, latency and cost per request, with the output check.
  • Re-tuned when you ship a new checkpoint.

Start a custom model

Tell us what you serve today. We reply within one business day with what we'd change and the gain we expect.

Questions

What does Inferoptimal charge?

Nothing for endpoints in your AWS account. AWS bills the instance to your account, and your credits apply. Inferoptimal cloud and custom model tuning are paid.

My AWS account has no GPU quota.

New AWS accounts start at 0 vCPUs for GPU instances. After you connect, the table shows your quota per region and files the increase for you; AWS usually answers within a day, and H100s can take longer. Meanwhile the Qwen3-0.6B starter runs on the CPU quota every account has, so you can launch right away.

Where does my data go?

Requests travel over HTTPS through Cloudflare to the model running in your AWS account. Inferoptimal does not store prompts or completions.

What can the connection do in my account?

See which instance types and quotas you have, file quota requests, and launch, stop and terminate instances it tags velobyte=1. Nothing else: no data, no other instances, no IAM. Delete the velobyte-… stack in CloudFormation to remove it.

Also available

Shared API

GLM 5.3 Flash on a pay-per-token API with prepaid credit. Change the base URL and go. This tier may use prompts for training; read the terms and data policy first.

Checking availability…

Open your account ↗
POST /chat/completions
curl https://inferoptimal.com/inference/v1/chat/completions \
  -H "Authorization: Bearer $VELOBYTE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "glm-5.3-flash",
       "messages": [{"role": "user", "content": "Hello"}]}'