pypi / ultralytics-platform
ultralytics-platform
What changed in ultralytics-platform 0.1, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-19.
4 breaking14 new APIs and options1 performance5 behavior and defaults4 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli initultralytics-platform 0.128 changes
Breaking changes 4
The API contract lowers the maximum export image size from 8192 to 4096.
In
ul cloud train,predict, andexport, a username before the first slash inproject=selects the workspace, and the remaining path becomes the project name.Using another workspace requires editor access; if the workspace exists but the user lacks access, the request fails before files are uploaded.
Failed or cancelled cloud jobs now exit with code 1; local input errors continue to exit with code 2.
New APIs and options 14
A new option lets similar-image searches include the source dataset and datasets containing copies of the query image. Copies of the query image itself remain excluded.
Sync and async image prediction methods accept up to 200 same-image seed labels—boxes, segments, or oriented boxes—for YOLOE26X prediction. Predictions can include confidence scores and use seed labels’ dataset class IDs; batch annotation does not support visual prompts.
Added an optional field to training callback responses and the Python TrainingMetricsResponse type, linking to the registered model’s page on the Platform.
The API and Python SDK expose task type, thumbnail URL, image count, and dataset size for bulk-import candidates; task type and dataset size are required.
The task field for Roboflow bulk-import candidates supports the listed YOLO task types.
Replaces gemini-3.1-pro-preview in the image-prediction model options.
Synchronous and asynchronous dataset searches now accept optional minimum and maximum image-count filters to narrow results.
Added muse-wonder-1.0 as an image-prediction option in the Platform API contract and synchronous and asynchronous Python SDK methods.
Accepts capture_dataset_version to save an immutable dataset version for a training run.
Reconnects to an existing export by export_id and downloads its result; Ctrl-C shows a resume command while the export continues remotely.
Accepts image or video URLs with a recognizable file extension, downloads them to the working directory, and processes them like local sources.
The Platform API and Python SDK support an optional comma-separated dataset license filter in synchronous and asynchronous Explore search methods.
Added six AMD Xilinx Versal AI Edge Series Gen 2 device targets for export, broadening deployment options for compatible hardware.
Added xilinx as a supported model export format in the API contract and Python SDK, enabling Python SDK users to request Xilinx-targeted exports with typed SDK support.
Performance 1
Already-compressed image formats such as JPEG, PNG, and WebP are stored without recompression, reducing CPU time but potentially making archives slightly larger.
Behavior and default changes 5
Replaces claude-haiku-4-5-20251001 in the image-prediction model list.
Relevance sorting is now available and is the default for searches; other searches continue to list newest results first.
Displays the GPU when one is reported by the Platform.
Prints per-image YOLO-style prediction lines instead of raw JSON; saved runs still include results.json.
Looks up or creates the project before packaging a local dataset, reporting Platform access problems sooner; a project may be created before local dataset validation completes.
Fixes and other changes 4
The API explains that reaching the detection cap may mean more objects exist, and that truncated generative output can omit objects or classes.
Project model listings exclude validation plots; retrieve an individual model to access them.
Malformed model and data addresses are rejected with a clearer error before uploads or Platform requests.
Workflow reruns no longer send duplicate Slack notifications.