inference Releases

Feature release adding fisheye camera calibration support and a heatmap visualization block for detections and tracklets. Includes various performance optimizations and maintenance improvements.

Feb 13, 2026

MNN 3.4.0 focuses on deepening GPU/QNN capabilities, optimizing attention computation for long texts, and improving GPU runtime stability. Key additions include Vulkan LLM support, Flash Attention on CPU and Metal, and a new unified `attention_mode` configuration.

Feb 7, 2026

Release candidate previewing the new inference-models prediction engine that will become the foundation for future development. Introduces faster model loading, TensorRT support, and multi-backend support (onnx, torch, trt).

Feb 6, 2026

Maintenance release adding new model support (RF-DETR NAS), enhanced VLM handling, improved error logging, and new AI model integrations including Claude Opus 4.6 and Gemini tool code execution.

Feb 6, 2026

This release expands the model library by adding large RF-DETRs, Seg-COCO models, and initial YOLO26 support with configurable API timeouts.

Jan 30, 2026

This update enhances OCR processing with new algorithms and fixes rotation metadata for video previews.

Jan 23, 2026

This release addresses Python deprecation warnings and improves video handling by adding new model aliases and cache mechanisms.

Jan 22, 2026

A maintenance release fixing postprocessing errors for YOLO26 and RF-DETR instance segmentation.

Jan 16, 2026

A minor release adding additional YOLO26 model aliases for better discoverability.

Jan 15, 2026

An early release introducing basic YOLO26 model aliases.

Jan 15, 2026

A major release introducing YOLO26 support and new Vision Language Models (Gemini 3 Flash and Claude Opus 4.5) for workflows.

Jan 15, 2026

This release introduces new workflow blocks for tracking object lifecycles and displaying text on images.

Jan 9, 2026

This release introduces support for advanced vision models like DepthAnythingV3 and Qwen3vl within Workflows. It enhances workflow capabilities with a new detections rollup block and expands Gemini integration with API key passthrough and support for newer models like Gemini 3 Pro. The update also includes various maintenance fixes to transformers and watchdog components.

Jan 6, 2026

Roboflow Inference adds Managed API Key support for Twilio SMS and MSS blocks, alongside a new v2 Camera Focus block utilizing Tenengrad measures. The release also enhances WebRTC capabilities by adding MJPEG support to the worker and implementing ACK-based flow control for batch processing. Several maintenance fixes address modal connections and visualization issues.

Jan 3, 2026

This release focuses on expanding the experimental inference module with SAM2 and SAM3D models while improving documentation for WebRTC streaming. It includes critical fixes for OWLv2 serialization and deserialization, along with a typo correction in ResNet model references. The update ensures stability for existing features while introducing new segmentation capabilities.

Dec 23, 2025

Workflows gain support for GPT-5.2 and a new classical OpenCV Motion Detection block, alongside new ResNet classifier aliases for easier model selection. The release includes maintenance updates for Jetson documentation, workflow testing, and metadata handling. It also adds several experimental models like OWLv2 and RF instant to the `inference-exp` module.

Dec 19, 2025

This release introduces streaming capabilities for Vision Language Models (VLMs) and a new SAM3 v3 block supporting RLE masks. It improves Jetson environment specifications and adds new parameters to the WebRTC SDK for better region and plan configuration. Various fixes address video track handling and workflow builder issues.

Dec 12, 2025

The update significantly expands the `inference-exp` module with new OCR models like DocTR, EasyOCR, and TROcr, alongside GroundingDino support. It introduces new versions of Claude and OpenAI workflow blocks and adds Roboflow Managed API Key support for Google Vision OCR. The release also includes a complete WebRTC SDK implementation and various maintenance fixes for Modal and builds.

Dec 10, 2025

This release adds official support for Jetpack 6.2 and introduces batch processing capabilities via cloud storage. It also includes conditional model preloading and an updated Gemini workflow block. A fix ensures that the lower bound confidence threshold is correctly applied.

Nov 26, 2025

A major release introducing Segment Anything 3 (SAM3) to the core inference module, enabling users to use the latest Meta AI segmentation model directly. The release provides comprehensive documentation and blog posts to help users leverage SAM3 for workflows, labeling, and rapid model building. This marks a significant upgrade to the segmentation capabilities of the platform.

Nov 19, 2025