July 31, 2026

Assets — one home for datasets, weights, and job outputs, with imports, your own buckets, and one-click attachment to Pods. Task storage now grows on demand.

Our largest release so far. Assets brings your datasets, model weights, LoRAs, and job outputs into one place in the workspace, and connects them directly to the Pods and tasks that use them. Task storage also stops being a fixed allowance — it now grows as your job writes.

Major features

Assets — one home for your data

Workspace › Assets is a shared list for everything your work reads and writes: training sets, checkpoints, adapters, configs, and the outputs your jobs produce. Everyone in the workspace sees the same list, and the same asset can feed a Pod, a task, or a plain download.

Each asset is a named series with versions underneath it, so re-uploading or re-harvesting a newer copy adds v2 rather than overwriting what you had. Files carry badges for type and format so you can tell a safetensors checkpoint from a pickled one at a glance.

Five ways to get data in

Assets can be created from whichever direction suits the work:

Upload from the browser — whole folders, resumable, so a dropped connection doesn’t cost you the transfer.
Harvest from a Pod — mark the paths a Pod writes to, and its outputs land back in Assets when the run finishes.
Import from Hugging Face or CivitAI — pull a model straight in without downloading it to your laptop first. The import runs with your access token or API key, so your access and each model’s licence terms apply.
Register from your own bucket — point at objects you already have in S3, R2, MinIO, or GCS and use them as assets without copying anything. You can also register an entire prefix as a single reference, which suits datasets far too large to duplicate.
Task outputs — see Policy changes below.

Attach assets to Pods and tasks

Pod create and edit now have an Assets section where you choose what to mount and where. Templates can declare the paths they expect — a model directory, an input directory — and the picker fills those slots, so you are matching data to a named slot rather than guessing mount paths.

Running Pods can take new assets without a restart, and a browser for your own buckets is built into the picker so you can pull from external storage in the same flow.

Task storage grows on demand. /inputs and /outputs used to be a fixed allowance you could run out of mid-run. They are now backed by a volume that expands as your job writes, sized from your hardware selection. The Task detail and submit screens show the sizing that applies.
Six more official templates, and a Science category. The catalog is now 50 templates, including additions for scientific and data workloads and Blender. Several now start as a non-root user, which suits images that expect it.
A file explorer for large collections. Assets with many files get a browsable tree with filtering, instead of one long list.
Console polish. A collapsible sidebar with auto-hide, corrected error-banner placement, and a consistent look for tables and selection across the Asset screens.
Security hardening across platform endpoints, continuing the edge work from the previous release. No action is needed on your side.

Policy changes

Task outputs are now assets, and are billed as storage

Previously, serverless task outputs were deleted automatically once a retention window passed. They are now registered as assets and are never deleted automatically — they stay until you delete them, or until the unpaid-balance lifecycle removes them.

What this means for you:

Task outputs count toward asset storage, billed per GB·month by the hour on the bytes that exist. The live rate is on the storage gauge at the top of the Assets page and on the Pricing page.
Outputs produced from this release onward appear in Assets automatically, and earlier task outputs have already been brought across, so nothing you previously generated has been lost.
Assets registered from your own bucket are not billed by us and are never deleted by us.

What happens when a balance runs out

Asset storage has a defined lifecycle when a workspace balance is depleted: a grace period where everything keeps working, then a period where new data can’t be added but downloads, exports, and deletion stay open, and finally deletion of assets in managed storage. Notifications go to the billing contact at each stage, and paying the balance restores full access immediately.

Assets registered from your own bucket are never touched by this.

Editing a Pod while it was still being created could leave it in an inconsistent state — this is now refused with a clear message until creation finishes.
Changing the template during Pod edit no longer discards the asset configuration you had already set.
Resetting a custom template no longer hangs on a spinner.
Resources without a name could be deleted with a single click and no confirmation.
Deploy requests that named a specific node could fail with a server error and leave a stuck operation behind.
Long asset downloads no longer fail partway through when a session token expires.