New Release: Ultralytics v8.4.0

Discussion
, ,
1

:tada: Ultralytics v8.4.0 is live — introducing Ultralytics YOLO26 :rocket:

Quick overview: v8.4.0 delivers the Ultralytics YOLO26 release — a re-engineered, edge-first family with native end-to-end (NMS-free) inference, major CPU speedups, and upgrades across detect, segment, classify, pose, and OBB. This release also includes an ARM64 Docker fix to improve deployment experience on modern hardware. :puzzle_piece:

Ultralytics YOLO26 Comparison Plots
Ultralytics YOLO26 Comparison Plots3840×1680 244 KB

:glowing_star: New Features (Highlights)

:rocket: Ultralytics YOLO26: native end-to-end, NMS-free

Ultralytics YOLO26 arrives in v8.4.0 with a streamlined architecture designed for edge and low-power environments—built for simpler exports and faster inference.

Key highlights:

  • :high_voltage: NMS-free end-to-end inference (native design, reduced latency, simpler deployment)
  • :racing_car: Up to 43% faster CPU inference (edge-first optimization)
  • :brain: MuSGD optimizer (hybrid optimizer inspired by modern LLM training stability ideas)
  • :wrench: No DFL (streamlines export + boosts compatibility with low-power hardware)
  • :bullseye: Task-specific improvements
    • Segmentation: semantic loss + multi-scale protos
    • Pose: RLE for higher-precision keypoints
    • OBB: angle loss to improve boundary behavior
  • :globe_with_meridians: YOLOE-26 open-vocabulary: zero-shot detection using text or visual prompts

:wrench: Improvements

  • :puzzle_piece: Edge-first philosophy: removing bottlenecks like DFL and NMS to maximize performance on constrained devices
  • :toolbox: Unified & versatile: end-to-end support across detection, segmentation, classification, pose, and OBB
  • :chequered_flag: Next-gen training stability: MuSGD helps improve convergence behavior and training stability

:bug: Bug Fixes

  • :white_check_mark: ARM64 Dockerfile fix for smoother builds and deploys on ARM devices

:shuffle_tracks_button: What’s Changed (PRs)

:package: How to Try v8.4.0

Update in seconds:

  • pip install -U ultralytics

Quick sanity check (Python):

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.predict(source="https://ultralytics.com/images/bus.jpg")

You can also browse the full release details directly in the official v8.4.0 release page, and review the full diff in the v8.4.0 full changelog.

:speech_balloon: Feedback Welcome

Please try v8.4.0 on your workloads (especially CPU/edge deployments and exports), and share:

  • performance numbers (CPU/GPU, batch size, imgsz, export target)
  • export experiences (ONNX, TensorRT, CoreML, etc.)
  • any regressions or odd behaviors you spot

Your feedback helps the whole YOLO community push the ecosystem forward.