AI models

AI models

Run an open model on its own rented GPU behind an OpenAI-compatible URL, in a project environment.

AI models are open models served on rented GPUs, behind an OpenAI-compatible URL with their own API key. They live in a project environment next to your apps, and GPU time is paid from the workspace's prepaid GPU credit.

The model decides its machine: one GPU of the class it needs, or one machine with 2, 4 or 8 GPUs when no single GPU holds it. AI models may be switched off on your platform or not be part of your plan: the command then exits 3 and says which.

See the AI models guide for the dashboard flow, prices and wake behaviour.

stackblaze models list

Alias: models ls

Lists the environment's AI models: model, machine, region, status, copies ready/asked.

usage
stackblaze models list [options]

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models get

Shows an AI model: URL, status, machine, copies and sleep policy. While it starts, the step, percent and download speed.

usage
stackblaze models get <name> [options]

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models deploy

Alias: models create

Deploys an AI model: a catalog id or any Hugging Face id (owner/name). Before creating, prints what the model will run on, the price range of one copy per hour and your GPU credit; a model that cannot run exits 1 with the reason. The deploy is refused (exit 3) when your credit does not cover one hour at the model's highest price.

usage
stackblaze models deploy <model> [options]

--wait polls until the model is running, printing the step, percent and download speed as they change, and exits 1 when the model failed or is not running within --timeout (default 1800). With --json the result is one document with the wait under wait.

A model put to sleep with models sleep does not wake on requests; on-demand models sleep by themselves after the sleep-after window.

FlagDescription
--name <name>Name of the model in the environment (default: a generated one)
--region <id>Where its GPUs are rented, from models sizes (default: the default region)
--copies <n>How many machines run the model and share the requests (plan-limited, default 1)
--gpu-type <name>Pin every machine to one GPU type from models sizes (default: any GPU of the class it needs)
--always-onNever sleep (billed while it runs)
--sleep-after <seconds>On demand: sleep after this many seconds without requests (300, 900 or 3600)
--waitPoll until the model is running
--timeout <seconds>Max --wait time (default 1800)

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

Examples

terminal
stackblaze models deploy Qwen/Qwen3-8B --wait
stackblaze models deploy Qwen/Qwen3-8B --name chat --region eu --copies 2
stackblaze models deploy Qwen/Qwen3-8B --always-on --gpu-type H200

stackblaze models check

Looks a model up without deploying: the machine it needs, whether it is gated, and the price — or why it cannot run (exit 1).

usage
stackblaze models check <model> [options]
FlagDescription
--region <id>Price it for this region
--gpu-type <name>Price it pinned to this GPU type

Also accepts the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models sizes

Alias: models options

GPU classes and price ranges, regions, GPU types you can pin, catalog models and your plan's limits. Prices are ranges per hour: from the cheapest machine for rent now up to the most a machine of that class can cost.

usage
stackblaze models sizes [options]

Also accepts the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models sleep

Puts an AI model to sleep now (no GPU runs; the model stays on disk).

usage
stackblaze models sleep <name> [options]

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models wake

Wakes a sleeping AI model (about a minute).

usage
stackblaze models wake <name> [options]

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models set

Changes when the model sleeps: --sleep-after <seconds> (on demand) or --always-on.

usage
stackblaze models set <name> [options]
FlagDescription
--sleep-after <seconds>Sleep after this many seconds without requests (300, 900 or 3600)
--always-onNever sleep (billed while it runs)

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models scale

Changes how many copies (machines) run the model and share the requests (plan-limited).

usage
stackblaze models scale <name> --copies <n> [options]
FlagDescription
--copies <n>Copies of the model

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models rename

Renames an AI model. Its URL and API key do not change.

usage
stackblaze models rename <name> <new-name> [options]

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models replace-gpu

Releases the model's GPU and rents a fresh one: same URL and key, a cold start follows. Asks you to type the model's name to confirm.

usage
stackblaze models replace-gpu <name> [options]
FlagDescription
-y, --yesSkip the confirmation prompt

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models logs

An AI model's logs: its lifecycle events, or the model server's own output.

usage
stackblaze models logs <name> [options]
FlagDescription
--source <source>events (rented, downloading, asleep…) or model (the model server)

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models metrics

Requests, response time, tokens, errors and GPU load of an AI model over a window.

usage
stackblaze models metrics <name> [options]
FlagDescription
--since <window>Window from 5m to 7d, e.g. 1h, 24h or 7d (default 1h)

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models url

Prints the model's OpenAI-compatible base URL (append /chat/completions).

usage
stackblaze models url <name> [options]

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models key

Prints the model's API key (stdout; --json: the key field). No other command prints it. The key is a credential: the server records who read it.

usage
stackblaze models key <name> [options]

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags

stackblaze models delete

Alias: models rm

Deletes an AI model and releases its GPU. Asks you to type the model's name to confirm.

usage
stackblaze models delete <name> [options]
FlagDescription
-y, --yesSkip the confirmation prompt

Also accepts the scope flags (-p, --project, -e, --env, -a, --app) and the output flags (-o, --output, --json), plus --token, --api-url and --org. Common flags