Skip to main content
The Endpoints screen lists your running models: “Running models you can call. Each one bills only while it is running.” From here you deploy base models, open an endpoint to manage it, and reach Aliases.
Valkyrie Endpoints screen

Endpoints: every deployment, with its URL, type, and status.

The endpoints list

Tabs filter by Active, All, Failed, and Terminated, each with a count. Search by name to find an endpoint across every page, and click a column header to sort your whole list. Each row shows: The buttons at the top right open Aliases, Pools, and Deploy a base model.

Deployment stages

An endpoint is only ready to use once it reaches Running. While it starts, it passes through:
Preparing means the machine is up and the runtime and model files are being set up. The screen shows this progression live, so you can see exactly where a starting endpoint is.
Large models can take a while to start. If an endpoint is still starting after 25 minutes, the dashboard shows Taking longer than usual. If the machine never becomes reachable, Valkyrie stops it automatically about 20 minutes after it comes up, with an error that explains why, so you are not billed for a machine you cannot use.
Besides the start-up stages, an endpoint can be Paused, Off (schedule) (with when it starts again, for example Off (schedule) · starts Tue 09:00), Stopped (idle), Stopped (budget), Interrupted (outbid), Terminated, or Failed. Terminated is only for an endpoint that does not come back by itself. See Deployment status and failures for what each means and what to do.

Endpoint detail

Click an endpoint to open it. The page has three tabs: Overview, Webhooks, and Access.
Valkyrie endpoint detail screen

An endpoint's Overview tab, with its URL ready to copy.

Overview shows the status pill and a Base model or Fine-tuned pill, the endpoint’s schedule in one line (see Schedules), when the endpoint was created, and the GPU it runs on. Model weights reads Cached when the last start reused the weights already on the machine (with their size), or Downloaded at start (with the size and how long the download took). It doesn’t show when Valkyrie didn’t measure it, for example on older endpoints. The Endpoint section has:
  • The Endpoint URL with a Copy button, while the model is running. It shows as soon as the machine is up, before the model has finished loading. While the endpoint is not running, the URL is hidden and a note says why and, for an endpoint off on its schedule, when it starts again.
  • The alias pointing at it, or a Create one link if there is none.
  • Create a key for this endpoint, for anyone who may use it. See Create a key for an endpoint.
  • The URL slug, which people who manage the endpoint can change with Rename URL.
Below that, the page shows:
  • What the machine is doing while it starts, with download progress and a boot timeline, and for a failed endpoint the failure and its Next step. See What the machine is doing.
  • A cost notice: the hourly price while running, or for a paused endpoint what its disk costs per day and when the pause ends.
  • The Diagnosis card, which puts the latest reports and the stored log tail in one place. See Diagnose a deployment.
  • The Pay as you go card (hidden while always on), the Always on card, and Schedules.
The actions for the endpoint are in its menu.

Operate an endpoint

A scheduled stop releases the machine, and a scheduled start rents a fresh one. While it is off, the endpoint reads Off (schedule) with the time it starts again, not Terminated.

Schedules

Each endpoint says when it runs, in plain words, at the top of its page and under its status on the Endpoints list: The exact start and stop times live under Schedules in the page’s Advanced section, with the next start and stop.

Create a key for an endpoint

Anyone who may use the endpoint (its owner, workspace members with use access, and people it is shared with) can create a key for it from its page, while it runs or while it is paused or off on its schedule:
1

Open the endpoint

Click Create a key for this endpoint in the Endpoint section and give the key a name.
2

Copy the key and the config

The key is shown once, with a ready-to-paste config for OpenCode, environment variables for any OpenAI-compatible client, and a curl test. Each carries the endpoint’s base URL (its alias when it has one), the key and the model name.
The key works only on this endpoint. Revoke it on the API keys screen. Requests made while the endpoint is off fail until it runs again.
A running endpoint holds a GPU and costs money even when idle. An endpoint that serves no requests for 2 hours, or for its own Pay as you go timeout, is stopped automatically unless it’s always on. Use stop, pause, or a schedule to control spend in between, see Wallet & billing.

Share an endpoint

On the Access tab, people who manage the model give members, teams, or email addresses Can use or Can manage access. Removing someone revokes the keys they hold for that model. Endpoints shared with you are marked in the list and when you create a key. See Share a model.

Inference pools

Pools on the Endpoints screen opens Inference pools: one endpoint over several deployments of the same model. The page is read-only. For each pool it shows whether it’s Active or Disabled, the model, how many request slots are in use, and its members with their state (Healthy, Unhealthy, Replacing, Starting, Draining), health, slots, and whether the pool provisioned them or they were added. A managed pool also shows its target size and its serving, provisioning, draining, and failed members. Create and change pools with the API or the MCP tools. To give a pool a memorable URL, point an alias at it. A pool alias needs an account-wide API key.

Deploy a base model

A base model is served as it is, with no fine-tune. Open Deploy a base model from Endpoints, from Home (“No fine-tune needed? Deploy a base model as it is”), or from the Models list.
Valkyrie Deploy a base model screen

Deploy a base model: pick a preset or enter any Hugging Face model path.

1

Choose the model

Search the presets by name or Hugging Face path. Each preset shows the GPU type it needs, its memory, and tags such as tool calling. Or enter any Custom HuggingFace model path.
2

Pick the hardware

Under Hardware, pick Small, Medium, or Large. Each shows its price range per hour, and Valkyrie rents the cheapest card in that size that holds the model. A size that can’t hold it reads Too small for this model. Choose a specific card opens the full card list and Number of cards. See Hardware sizes.
3

Configure and deploy

Review the configuration: the cost estimate, the serving engine, the disk, and when it should run. Add an optional alias, and deploy. You land on the new endpoint’s page, where you can follow its stages to Running. See Deploy a model for each setting.

Deploy a fine-tuned version

When a version on Models has completed, click Deploy to open its deploy page.
1

Pick the serving settings

Under Serving settings, pick the Hardware: Small, Medium, or Large, each with its price range per hour. A fine-tuned version needs at least 24 GB of memory per card, so a size too small for it reads Too small for this model. The page opens on the smallest size that holds it, and Choose a specific card opens the full card list. See Hardware sizes. Choose the Engine: vLLM (recommended) or Ollama. The Advanced section lets you set the context length, memory use, and an optional cluster name.
2

Choose when it runs

Under When should it run?, pick Office hours (starts and stops by itself on weekdays, between the hours you choose, in your timezone), Pay as you go (stops after the quiet period you choose under Stop after, 30 minutes to 24 hours), or Until stopped (runs until you stop it, or until it serves no requests for 2 hours). Tick Keep always on to keep it running 24/7 instead. Each choice shows its approximate monthly cost at the current hourly price. The same choice appears when you deploy a base model. See Always-on and idle stop.
3

Check the summary

The Summary shows the model, the chosen card, the engine, and Cost while running per hour. Billed from your wallet while the endpoint runs. Stop it any time from its page.
4

Deploy

Click Deploy endpoint. You land on the new endpoint’s page.
Once a version is deployed, its page on Models links to the running endpoint instead of showing Deploy.

Call the endpoint

A ready endpoint exposes an OpenAI-compatible API at its URL (/{deployment_slug}/v1/...). Attach an alias for a memorable URL, try it first in the Playground, and see Run inference for request examples.
A fine-tuned endpoint serves its model as fine-tuned-<job id>. You do not need to match that name exactly: Valkyrie routes requests sent to the endpoint’s URL to your fine-tuned model whatever model name the client sends.