π Summary
Ultralytics YOLO26 has arrived. Re-engineered from the ground up by @glenn-jocher, @Laughing-q, and the Ultralytics YOLO team, YOLO26 is purpose-built for edge and low-power environments. This release introduces a streamlined, native end-to-end NMS-free architecture, delivering faster, lighter, and more accessible deployment across all platforms.
π Whatβs New
- π NMS-Free End-to-End Inference: A fully native design that eliminates the need for Non-Maximum Suppression (NMS) post-processing, significantly reducing latency and simplifying export complexity.
- β‘ 43% Faster CPU Inference: Optimized specifically for edge computing, achieving real-time performance on CPU-only devices.
- π§ MuSGD Optimizer: A pioneering hybrid optimizer combining SGD with Muon. Inspired by Kimi K2 (Moonshot AI), this brings LLM-grade optimization stability to computer vision training.
- π§ Streamlined Architecture (No DFL): Complete removal of Distribution Focal Loss (DFL) to streamline model export and maximize compatibility with low-power hardware.
- π― Task-Specific Enhancements:
- Segmentation: Added semantic loss and multi-scale protos.
- Pose: Implemented RLE for high-precision keypoints.
- OBB: Introduced angle loss to resolve boundary discontinuities.
- π YOLOE-26 Open-Vocabulary: Zero-shot inference capabilities allow for the detection of any object class using text or visual prompts.
π― Impact & Vision
- Edge-First Philosophy: By removing architectural bottlenecks like DFL and NMS, YOLO26 maximizes speed on resource-constrained hardware.
- Unified & Versatile: Full end-to-end support for detection, segmentation, classification, pose, and OBB.
- Next-Gen Training: The MuSGD optimizer bridges the gap between LLMs and Vision, offering faster convergence and stable training.
What's Changed
- Fix ARM64 Dockerfile by @glenn-jocher in #23175
ultralytics 8.4.0YOLO26 Models Release by @glenn-jocher in #23176
Full Changelog: v8.3.253...v8.4.0
