npm / replicate
replicate
What changed in replicate 1, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-28.
3 breaking1 deprecations33 new APIs and options3 behavior and defaults1 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initreplicate 141 changes
Breaking changes 3
In sync mode, output.url() may contain a data-uri instead of a remote URL; pass wait: { mode: "poll" } to replicate.run() if a URL is required.
For models that output files, replicate.run() now returns FileObjects rather than URL strings by default.
Pass useFileOutput: false to the Replicate constructor to opt out of FileOutput and revert to the previous behavior.
Deprecations 1
The prediction request field is deprecated because supported models always include a stream entry in the returned prediction URLs.
New APIs and options 33
The useFileOutput option can be set to false to output the HTTP or data URL directly.
The request header sets the maximum prediction runtime before automatic cancellation; durations support seconds, minutes, and hours, with a minimum of 5 seconds.
Prediction responses expose the absolute time at which a running prediction will be automatically canceled if it has not completed.
Create a deployment by specifying its name, model, version, hardware, and minimum and maximum instances.
Delete a deployment if it has been offline and unused for at least 15 minutes.
Retrieve a deployment by name, including its current release.
Create a prediction using a deployment;
Prefer: waitwaits up to 60 seconds before returning a prediction that may still be starting.Update deployment hardware, minimum or maximum instances, or its underlying model version; each update increments the current release number.
Upload a file with optional JSON metadata; file content and a filename are required, and files must be less than 100MB.
Delete a file; subsequent requests to the deleted file resource return 404 Not found.
Download a file using its owner, an expiration timestamp, and an HMAC-SHA256 signature.
Get the details of a file by its ID.
List files created by the user or organization associated with the API token, with pagination.
List available hardware SKUs for running models.
Create a model by providing its owner, name, visibility, and hardware.
Delete a model you own if it is private and has no associated versions.
List example predictions saved by a model author, so developers can retrieve a model’s illustrative predictions.
Get a model and its latest version's input and output schema.
List public models with pagination and optional sorting by model creation time or latest-version creation time.
Choose ascending or descending model-list sort order; descending is the default.
Create predictions directly using official models.
Retrieve a model’s README content as plain-text Markdown.
Search public models using a text query, returning paginated model results.
Update a model's description, README, GitHub URL, paper URL, weights URL, or license URL.
Request deletion of an eligible private model version and its associated predictions and output files.
Filter listed predictions to those created at or after a specified ISO 8601 date-time.
Filter listed predictions to those created before a specified ISO 8601 date-time.
Set the request header to wait for model output, using wait=n for a value from 1 to 60 seconds.
Search public models, collections, and documentation; model results include generated descriptions, tags, and relevance scores.
A training can be retrieved by ID, with status, output or error details, and metrics for terminated trainings.
A training can be canceled by its ID.
The default webhook endpoint's signing secret can be retrieved to verify that webhook requests are from Replicate.
FileObject implements ReadableStream, so developers can access its URL, write it to disk, stream it, or read it in chunks without an additional fetch().
Behavior and default changes 3
For models that output files, event.data is a FileOutput instance by default and can be written to disk.
Prediction and training statuses include aborted for work terminated before it starts running, such as when a deadline is reached before startup.
Filter predictions by creation source; filtering for web predictions limits results to the last 14 days.
Fixes and other changes 1
Types are updated to match the current HTTP API.