Skip to main content
A short tour of the objects you’ll work with in Valkyrie and how they relate.

Models, deployments, and jobs

Fine-tuning job

A training run that adapts an open base model to your dataset using a managed Unsloth pipeline. Produces model outputs (weights/adapters) stored for you.

Deployment

A running, servable model behind an OpenAI-compatible endpoint. It can serve a fine-tuned model or any Hugging Face base model, using vLLM or Ollama.
A typical lifecycle is dataset → fine-tuning job → deployment → API calls, but you can skip straight to a deployment if you just want to serve an existing model.

GPUs are provisioned on demand

Valkyrie doesn’t keep idle GPUs around. When you launch a job or a deployment, it provisions a GPU from a cloud marketplace, runs your workload, and releases it when it’s no longer needed. Deployments can auto-stop when idle and be resumed on demand, so you pay for compute you actually use.
You never manage GPU machines, drivers, or serving infrastructure yourself. Valkyrie handles provisioning, setup, health monitoring, and teardown.

The wallet

Billing is prepaid and per-use. You add credit to a wallet, and jobs and deployments draw against it. Your balance is computed from the full ledger of credits and debits, so every charge and top-up is auditable. See Quota & billing.

Aliases

An alias is a stable, human-readable name for a deployment. Instead of calling a raw deployment slug, you call:
Aliases let you swap the underlying deployment (for example, after redeploying a new version) without changing the URL your applications use. See Aliases.

Ways to use Valkyrie

Dashboard

The web app for people: manage keys, datasets, jobs, deployments, and spend.

OpenAI-compatible API

Call your deployed models from any OpenAI-compatible client or SDK.

MCP server

Let an AI agent provision, deploy, and operate models programmatically.

Authentication at a glance

API keys are prefixed by purpose, for example, deployment keys begin with vk_dep_ and MCP keys begin with vk_mcp_. See Authentication.