Vastbeam Request an endpoint

Vastbeam Inference

Your model. Your endpoint. Your environment.

Vastbeam Inference serves open-source and fine-tuned models on dedicated endpoints, on our infrastructure, in your cloud, or on your hardware, with the audit, guardrails, and evidence that regulated teams need to get it approved.

Request an endpoint Talk to us about a private installation
Drop-in
Speaks the OpenAI and Anthropic APIs. Change the base URL and the key.
Dedicated
Your capacity. No other tenant on your endpoint.
Nothing kept
Every request is recorded. No prompt or output is.

Where it runs

Run it where your data must stay

Choose where your model runs.

  1. 01

    On Vastbeam

    We provision and operate the endpoint. You get a URL and a key.

  2. 02

    In your cloud

    We install and operate Inference inside your own cloud account, in the region you choose.

  3. 03

    On your hardware

    We install Inference in your data center, and your data never leaves it.

Inference runs on NVIDIA, AMD, Intel, and Apple GPUs, on accelerators from AWS, Google, Huawei, IBM, Baidu, and Cambricon, and on x86 and ARM CPUs.

Wherever it runs, your prompts and outputs stay inside the environment you chose. That is what teams in healthcare, financial services, and government need, and what any team with a residency or privacy obligation gets.

Sensitive data

Nothing sensitive is kept

Inference never stores or logs a prompt or an output. This is not a setting someone can get wrong: a logging policy that would keep content is refused.

Every request leaves an audit record: who called, which model, when, and how many tokens. The record holds nothing of what was said, not even a hash of it. It is append-only, kept for the period you set, and refused requests are recorded with the reason.

Sign-in and roles come from your identity provider. Workspaces separate teams, models, and policies, and each workspace decides what its users may call.

Guardrails run inside your environment, on every request. Set each workspace to detect, scrub, or block personal data, and prompt-injection attempts are blocked in every mode. If a guardrail is unavailable, requests are refused, never let through unchecked.

Approval

Ready for security and compliance approval

Built in and always on, so approval is a review, not a project.

A private AI deployment is approved by people who never call the API: security, compliance, and procurement. Inference gives them what they ask for, without a slide deck.

Before install, a preflight check names every blocker and its fix. After install, readiness checks report the state of the deployment, in the same terms, for as long as it runs.

The Evidence Pack is a generated bundle for reviewers.

It covers

  • the deployment and its data flow
  • what is and is not logged
  • identity and roles
  • retention
  • an audit sample
  • readiness
  • the model inventory
  • cost
  • backup and restore

Names are left out by default, so it is safe to hand to a reviewer outside your team. It maps each fact to the HIPAA Security Rule and SOC 2 control families as a review aid, not a certification.

The endpoint

Any model, your endpoint

Pick an open-source model, or bring your own fine-tuned weights. Vastbeam Inference serves it on an endpoint that is yours alone: your capacity, your throughput, no other tenant on it.

The endpoint speaks the APIs your code already uses: OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages, with streaming, tools, structured output, and vision. Code written against Vastbeam Gateway runs against Inference with a change of base URL.

import os
from openai import OpenAI

client = OpenAI(base_url="https://<your-endpoint>/v1", api_key=os.environ["VASTBEAM_API_KEY"])
reply = client.chat.completions.create(model="<your-model>", messages=[{"role": "user", "content": "Hello"}])

Inference is for teams that need a model of their own, at a capacity they control, in a place they choose.

Throughput

Built for throughput

A dedicated endpoint is only worth having if it is fast and full. Inference is built to serve more requests from every GPU, to keep the first token fast when the endpoint is busy, and to bring the cost of each token down as load goes up.

Add capacity to an endpoint and the URL does not change. Your application never knows.

It is built by a team that has run inference at production scale, and tuned for the models that teams run themselves.

Cost

Know what it costs

See tokens and cost by workspace and by model, updated as requests run. Put your own hosting cost beside the cloud per-token rate for the same model, and see what running it yourself saves.

Set quotas per workspace or per model, in requests or tokens, per minute, per day, or per month. A request over quota is refused before the model runs.

Export usage and cost as CSV for finance.

Start

Request an endpoint

Vastbeam Inference is in pilot. Tell us what you want to run and where, and we will provision an endpoint for your team by hand.

We reply to this address.

The country or region your data must stay in.

Model names, or tell us you have your own fine-tuned weights.

Expected load, timing, and anything your reviewers will ask.

We use these details only to answer your request. Read the privacy notice.

Running it inside your own cloud account or data center? Talk to us about a private installation.

Want pay-per-token access instead? Write to us about Vastbeam Gateway.