What's Changed

• Serialize parent origin metadata for crops/slices by @brunopicinin in #1672
• Update perspective correction for dimensionality by @lou-roboflow in #1674
• "USE_INFERENCE_EXP_MODELS" by @hansent in #1657
• Add support for ingesting and processing rtmp stream remotely (on modal or process) and stream back processed frames over webrtc by @grzegorz-roboflow in #1675
• Add GitHub Actions workflow for Jetson 6.2.0 builds by @alexnorell in #1671
• uv lock inference-experimental by @hansent in #1677
• Add rtsp placeholder to webrtc_worker by @grzegorz-roboflow in #1683
• Set HF_HOME when running webrtc_worker on Modal by @grzegorz-roboflow in #1682
• Make webrtc_worker time out gracefully if WEBRTC_MODAL_FUNCTION_TIME_LIMIT is set by @grzegorz-roboflow in #1684
• add file lock for downloading core models by @hansent in #1680
• TRT Compilation of models in inference-cli by @PawelPeczek-Roboflow in #1679
• Remove (currently broken) trt-*-cu13 dependencies from inference-exp by @PawelPeczek-Roboflow in #1689
• Handle bursty rtsp by @grzegorz-roboflow in #1691
• Optimize bytes_to_unicode by @grzegorz-roboflow in #1676
• Drain rtsp queue by @grzegorz-roboflow in #1693
• InferenceExpObjectDetectionModelAdapter: dont optimize_for_inference by @hansent in #1694

Full Changelog: v0.59.1...v0.60.0

💪 Added

📺 hosted webrtc

Thanks to @grzegorz-roboflow we have a great news 🔥

With this release of inference it is now possible to easily process video streams over webrtc with inference!

We prepared example script to showcase how to process streams from your python scripts. Start inference server, and then stream web-cam in order to start processing the stream and to see live preview produced by inference:

python examples/webrtc/webrtc_worker.py \
    --workspace-id="<your workspace>" \
    --workflow-id="<your workflow id>" \
    --api-key="<your API key>" \
    --inference-server-url="http://127.0.0.1:9001" \
    --realtime

Locally stored files can also be streamed and processed:

python examples/webrtc/webrtc_worker.py \
    --workspace-id="<your workspace>" \
    --workflow-id="<your workflow id>" \
    --api-key="<your API key>" \
    --inference-server-url="http://127.0.0.1:9001" \
    --source "/path/to/file.mp4" \
    --realtime

The processing can be performed on Modal simply export WEBRTC_MODAL_TOKEN_ID and WEBRTC_MODAL_TOKEN_SECRET with your credentials, and inference takes care of creating Modal app and processing stream on Modal function. All aspects of Modal execution are controlled through environmental variables. On-demand stream processing on CPU or GPU is extremely easy with this feature!

The feature is also available through our app https://app.roboflow.com:

Check out details #1612

Other new features

• Add RLE format support to SAM2 by @lrosemberg in #1651
• Seg preview workflow block by @hansent in #1647

🔧 Fixed

• Fix assertion in platform tests by @PawelPeczek-Roboflow in #1644
• Fix base detection docstrings default confidence and IOU by @dagleaves in #1660
• Fix ModelConfig incorrect confidence environment variable by @dagleaves in #1659
• Fix image dimensions in detections stitch block output by @brunopicinin in #1654
• Claude 3.5 is deprecated as of 28th of October 2025, replacing with Claude 4.5 by @grzegorz-roboflow in #1668

🏡 Maintenance

• Healthz and Readiness should not be authed by @bigbitbus in #1650

🥼 inference-experimental

• Add RFDetr Segmentation to inference-exp by @PawelPeczek-Roboflow in #1645
• Loose RFDetr Seg integration tests assertions for ONNX which seems to be quite numerically unstable by @PawelPeczek-Roboflow in #1646
• Fix assertions in RFDetr segmentation by @PawelPeczek-Roboflow in #1649
• Add fixes for TRT error handling by @PawelPeczek-Roboflow in #1665
• Bump version of inference-exp by @PawelPeczek-Roboflow in #1653
• TRT inplementation for more models by @PawelPeczek-Roboflow in #1666
• Fix issue with inference resolution not denoted correctly in RFDetr by @PawelPeczek-Roboflow in #1667

Full Changelog: v0.59.0...v0.59.1

💪 Added

🍾 Workflows Execution Engine v1.7.0

This release refines how Workflow step errors are reported when interacting with the Roboflow platform. Client-side issues such as invalid API keys or model IDs now raise ClientCausedStepExecutionError with accurate HTTP status codes (400–404) instead of a generic StepExecutionError with HTTP 500.

Motivation

The change ensures that client misconfigurations are clearly distinguished from server-side errors, improving observability and making debugging easier.

Compatibility

Most users should not be affected. However, if an integration relies on the legacy error type or status code, it can be temporarily restored by setting:

DEFAULT_WORKFLOWS_STEP_ERROR_HANDLER=legacy

This update brings more accurate error reporting with minimal impact on existing workflows. See details of the change in the Workflows change log and in PR #1640

🥼 Changes inference-experimental

• Add changes to inference-exp deeplab sem-seg model by @PawelPeczek-Roboflow in #1629
• Adjust the name of yolov5 weights by @PawelPeczek-Roboflow in #1633
• Add tests for Yolov7 implementation by @PawelPeczek-Roboflow in #1634
• Add ability for inference-exp to parse modelVariant from RFAPI by @PawelPeczek-Roboflow in #1641
• Add PaliGemma 1 to inference-exp by @PawelPeczek-Roboflow in #1639
• YOLACT in inference experimental by @PawelPeczek-Roboflow in #1637
• Adjust YOLACT tests assertions for GPU tests to pass by @PawelPeczek-Roboflow in #1638

Full Changelog: v0.58.3...v0.59.0

What's Changed

• Fix RFDetr issue with class remapping and ensure model cache paths sanitisation by @PawelPeczek-Roboflow in #1619
• Update Overlap block UI configuration by @brunopicinin in #1622
• Add changes to fix multi-class keypoints detection by @PawelPeczek-Roboflow in #1626
• Add implementation of yolov8-cls to inference-exp by @PawelPeczek-Roboflow in #1627
• bugfix: incorrect parsing of content-type breaks API_KEY authenticati… by @tonylampada in #1628
• Update size measurement video URL by @rfjames in #1625
• Fix JetPack 6.2 Dockerfile: Build ONNX Runtime from source with ARM64-compatible headers by @alexnorell in #1623
• Feat/include errors in usage by @grzegorz-roboflow in #1621
• Set workflow execution log level to debug by @bigbitbus in #1620
• bump version to 0.58.3 by @hansent in #1630
• Add Claude Haiku 4.5 to workflow block by @brunopicinin in #1632

New Contributors

• @rfjames made their first contribution in #1625

Full Changelog: v0.58.2...v0.58.3

🔧 Fixed

• Honor mask decode args for rf-detr seg model by @probicheaux in #1613
• Fix image preprocessing block to accept dynamic parameter references by @rf-elliot in #1616

🏡 Maintenance

• Report total remote execution time when running benchmarks by @grzegorz-roboflow in #1615
• add gcs/s3 example scripts by @japrescott in #1617
• Bump the version by @PawelPeczek-Roboflow in #1618

🏅 New Contributors

• @japrescott made their first contribution in #1617
• @rf-elliot made their first contribution in #1616

Full Changelog: v0.58.1...v0.58.2

🔧 Fixed

We've detected a bug that was in inference since 0.55.0 release - setting confidence to 0.0 was historically prevented globally in inference server. Since we wanted to remove that constraint, in 0.55.0 release, we make it possible to set such confidence threshold, but that turned out to lead to OOM issues with some models - namely instance segmentation ones, which in some cases may yield large amount of segmentation masks using wast amount of RAM. To provide quick-fix, we've added CONFIDENCE_LOWER_BOUND_OOM_PREVENTION env variable flag to control the minimal allowed confidence, set by default to 0.01.

Review the change here: #1611

🏡 Maintenance

• Add disable model monitoring on legacy requests by @rvirani1 in #1607
• Make triggering anchor optional in line_counter block by @grzegorz-roboflow in #1608

Full Changelog: v0.58.0...v0.58.1

💪 Added

🤯 RFDetr Seg — Fast, Accurate Segmentation in Roboflow 💥

• Train and deploy RF-DETR Seg models directly in Roboflow
• Enjoy state-of-the-art accuracy with latency
• Run models seamlessly through Roboflow inference or the rf-detr packages

This release pushes the boundary of fast, accurate segmentation — and it’s now available to explore in Roboflow. Check out our 📖 blog for more details and try out your RFDetr model in inference:

import os
import supervision as sv
from inference import get_model
from PIL import Image
from io import BytesIO
import requests

url = "https://media.roboflow.com/dog.jpeg"
image = Image.open(BytesIO(requests.get(url).content))

model = get_model("<your-rf-model-id>")
predictions = model.infer(image)[0]
detections = sv.Detections.from_inference(predictions)

🔌 Improved

• improve speed and memory usage of masks2poly and masks2multipoly by @hansent in #1586

Full Changelog: v0.57.4...v0.58.0

What's Changed

• Increase default rfdetr resolution limit by @grzegorz-roboflow in #1606
• Add query param to disable model monitoring by @rvirani1 in #1585

Full Changelog: v0.57.3...v0.57.4

What's Changed

• Add Claude Sonnet 4.5 to the Anthropic workflow block by @brunopicinin in #1601
• Fix issue with inference-exp Yolov8-pose TRT implementation by @PawelPeczek-Roboflow in #1602
• Limit rfdetr input resolution when loading through onnx by @grzegorz-roboflow in #1603

Full Changelog: v0.57.2...v0.57.3

💪 Added

Easy OCR 🤝 inference

Thanks to @lou-roboflow we have EasyOCR model supported in inference and Workflows ecosystem.

⚡ Improved

• ⚡️ Speed up function post_process_ocr_result by 765% in PR #1577 (New_OCR_Blocks) by @codeflash-ai[bot] in #1589
• Reduce latency resulting from VideoFrameProducer by @grzegorz-roboflow in #1579

🚧 Fixed

• Fix "Getting Started" python version requirement by @brunopicinin in #1581
• Fall back to 'Fit (grey edges) in' when 'Fit (reflect edges) in' / 'Fit within' / 'Fill (with center crop) in' is passed as resize method by @grzegorz-roboflow in #1584
• Update aiortc dependency by @cdeil in #1583
• Fix SyntaxWarning in core/devices/utils.py by @cdeil in #1582
• Guard owlv2 and transformers artifacts download with lock by @grzegorz-roboflow in #1578
• Fix test_rfdetr_base_with_background_class - by @grzegorz-roboflow in #1590
• Fix ocr changes by @PawelPeczek-Roboflow in #1592
• Fix issues with EasyOCR - cache location and requirements by @PawelPeczek-Roboflow in #1594
• Fix docker builds by @PawelPeczek-Roboflow in #1595
• Add bugfix for easyocr by @PawelPeczek-Roboflow in #1596
• Fix ocr again forth time by @PawelPeczek-Roboflow in #1597
• Pass execution_id header based on ContextVar set by http_api middleware by @grzegorz-roboflow in #1593
• Fix: Add psutil dependency to resolve slim Docker image startup failures by @antsa-alec in #1580
• Fix issues with EasyOCR and integration tests by @PawelPeczek-Roboflow in #1598
• Retain empty 'parent_id' field in OCR response by @PawelPeczek-Roboflow in #1599
• Fix bug in florence 2 old inference by @PawelPeczek-Roboflow in #1591

🥼 Changes inference-experimental

• support of Torch script backend for models
• support for models-specific features in auto-negotiation
• foundations for RF trained models support in inference-exp
  See details: #1450

🏅 New Contributors

• @antsa-alec made their first contribution in #1580

Full Changelog: v0.56.0...v0.57.2