New Release: Ultralytics v8.4.3

Discussion
, ,
1

:megaphone: Ultralytics v8.4.3 is out! :rocket:

TL;DR: Ultralytics v8.4.3 speeds up Ultralytics Platform NDJSON dataset conversion/downloads :high_voltage:, improves training metric correctness & reliability :brain:, and refreshes defaults/docs to make YOLO26 the smoothest starting point :books:.

:glowing_star: New Features

:rocket: Faster NDJSON → YOLO dataset conversion (Ultralytics Platform data)

Big speed + robustness wins for Platform-backed datasets in PR #23257 by @glenn-jocher:

  • Lazy-loads aiohttp only when NDJSON conversion is used (faster startup, fewer unnecessary deps) :package:
  • Simplifies async image download logic + improves concurrency scaling with dataset size :high_voltage:
  • More robust handling of Platform URL inputs :globe_with_meridians:
  • Includes version bump 8.4.2 → 8.4.3 :bookmark:

:globe_with_meridians: Configurable Ultralytics Platform base URL

For teams testing staging/local setups, PR #23256 by @glenn-jocher adds:

  • ULTRALYTICS_PLATFORM_URL env var to point callbacks/links to custom Platform environments :test_tube:

:chart_increasing: Improvements

:label: Defaults and examples move to YOLO26

Docs/examples now default to yolo26n.pt in PR #23242 by @Laughing-q:

  • YOLO() and CLI fallback model become yolo26n.pt :white_check_mark:
  • Helps new users start on the latest recommended model family (YOLO26) with fewer naming bumps

:gear: Optimizer + warmup logic is more reliable

Training behavior is more predictable in PR #23234 by @Laughing-q:

  • Parameter groups explicitly labeled; warmup LR targets the bias group by name (not position) :fire:
  • “Auto” optimizer strategy simplified around MuSGD with improved defaults :control_knobs:

:puzzle_piece: IMX inference/export consistency

Better shape-handling and a simpler decode path in PR #23235 by @Laughing-q:

  • More robust anchor/stride refresh for changing input shapes :wrench:

:bar_chart: Benchmark tables clarified (end-to-end metrics)

README tables now clearly distinguish e2e evaluation metrics in PR #23238 by @Laughing-q.

:strawberry: Raspberry Pi 5 guide refreshed with YOLO26 benchmarks

Updated benchmarks + ExecuTorch results in PR #23227 by @lakshanthad.

:bug: Bug Fixes

:person_standing: Pose training logs are more accurate

Avoids reporting rle_loss when unsupported in PR #23230 by @lmycross:

  • Cleaner logs, fewer confusing metrics :receipt:

:white_check_mark: Fix duplicated Results.summary() entries

Prevents duplicated rows in summaries in PR #23218 by @xusuyong:

  • Cleaner analytics/logging :broom:

:tada: New Contributors

Welcome @xusuyong — first contribution landed via PR #23218! :raising_hands:

:test_tube: Try it now

Upgrade with:

pip install -U ultralytics

Quick sanity check with YOLO26:

from ultralytics import YOLO

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

:magnifying_glass_tilted_right: Changelog & Release

Browse the full release details in the v8.4.3 GitHub Release, or review the complete diff in the v8.4.2 → v8.4.3 changelog comparison.

:speech_balloon: Feedback

Tried v8.4.3 yet? Let us know what improved (or what didn’t) — especially if you’re using Ultralytics Platform datasets, training at scale, or exporting IMX. Your reports help us keep YOLO fast, reliable, and easy to use.