New Release: Ultralytics v8.4.26

Discussion
, ,
1

:tada: Ultralytics v8.4.26 is here

Quick summary: v8.4.26 delivers a solid round of reliability and usability improvements across Ultralytics YOLO workflows, with smarter Ultralytics Platform dataset handling, more robust ul:// URI resolution, and an important FP16 SAM TinyViT inference crash fix. We’ve also polished CI and docs to make the overall experience smoother. :rocket:

If you’re working with Platform datasets, SAM models, or remote/cloud-based pipelines, this release should make things feel noticeably more stable and convenient.

:glowing_star: Highlights

:white_check_mark: Smarter Platform NDJSON dataset handling

With PR #23990 by @glenn-jocher, NDJSON datasets imported through Platform now handle missing validation/test splits more gracefully.

If a dataset includes train but no val or test, Ultralytics will now automatically create a small validation split from the training data instead of failing immediately.

Why this matters:

  • Less friction when converting or importing datasets :package:
  • Deterministic split behavior for consistency
  • Helpful warning so users can still create a proper manual validation split for best results

This is especially useful for fast-start workflows where dataset structure may not be perfect on first import.

:globe_with_meridians: More reliable Ultralytics Platform URI resolution

Also in PR #23990 by @glenn-jocher, ul://... resolution is now more robust.

Improvements include:

  • Retry logic for transient connection failures
  • Faster connect timeout plus longer read timeout for larger dataset operations
  • Clearer handling for auth, permission, not-found, and processing states

Impact:

  • Better reliability in cloud and remote workflows :cloud:
  • Fewer flaky failures during dataset access
  • Smoother experience when working with larger Platform-hosted assets

:brain: FP16 SAM TinyViT inference crash fix

With PR #23780 by @Edwin-Kevin, we fixed a half-precision inference issue affecting SAM TinyViT models such as mobile_sam.pt.

The root cause was a dtype mismatch in cached tensors during inference. The fix adjusts model setup order so cached eval-time tensors align correctly with FP16 expectations, and adds CUDA coverage with half=True.

Impact:

  • More stable FP16 segmentation inference on GPU :bullseye:
  • Fewer runtime dtype crashes
  • Better deployment behavior for SAM-based pipelines

A big thank-you to @Edwin-Kevin for the contribution—and congratulations on a first contribution to the repo! :raising_hands:

:books: Docs and developer experience

:movie_camera: Platform docs now include onboarding videos

PR #23986 by @RizwanMunawar adds embedded onboarding videos to key Ultralytics Platform documentation pages, helping new users get started faster with:

  • Quickstart
  • Account
  • Data
  • Train
  • Deploy

This should make onboarding much more accessible for users exploring annotation, training, deployment, and monitoring in Platform.

:speech_balloon: Docs chat update

PR #23989 by @glenn-jocher updates the docs chat script to v0.2.7, continuing ongoing polish across the docs experience.

:gear: CI and maintenance improvements

This release also includes cleanup and stability improvements to our maintenance pipeline, mainly through PR #23990 by @glenn-jocher:

  • Removed the dedicated HUB CI job and related manual trigger
  • Standardized many CI runners from cpu-latest to ubuntu-latest
  • Temporarily pinned GPU CI to torch<2.11 until CUDA 13 driver support is ready

These changes help reduce CI complexity and improve release consistency. :hammer_and_wrench:

:memo: What changed

:raising_hands: New contributor

A warm welcome to @Edwin-Kevin, who made their first contribution with PR #23780! :tada:

:magnifying_glass_tilted_left: Try it out

You can explore the release on the v8.4.26 release page or review everything included in the full changelog for v8.4.25…v8.4.26.

To upgrade locally:

pip install -U ultralytics

If you’re starting a new project, we recommend using YOLO26 on Ultralytics Platform, our latest stable model family, which is smaller, faster, more accurate than YOLO11, and natively end-to-end.

:speech_balloon: Feedback

Please give v8.4.26 a try and let us know how it works for your workflows—especially if you’re using Platform datasets, ul:// assets, or FP16 SAM inference. Your feedback helps the Ultralytics team and the broader YOLO community keep improving every release.