The New YOLO26 is Finally Released! Slower Results, but Intelligent!

Hello everyone, and I hope you are doing well.

After many months of development, YOLO26 was officially released today January 14th with all the models out and its availability for use using the Ultralytics Python package.

Let’s break it down, the results, and how it is compared to other technologies.

their claimed graph

YOLO26 Features

This is YOLO26, changing the naming (which technically this is 14) to show that it is going to be the supreme model of 2026. Developed by Ultralytics, these models are the new SOTA in the YOLO space featuring:

• Ultralytics Python Package Compatibility: This Python Package makes it super easy to train these models and one can use it immediately at release. Previous datasets built with this package are can seamlessly change from the previous model to YOLO26.
• Services with Ultralytics HUB: Ultralytics HUB is one of their apps and gives ease when working with datasets.
• GitHub Contribution: There is a GitHub repositiory that is ready for the flood of questions regarding YOLO and support. The GitHub will also be ready for contributions by developers.

They were able to achieve:

• DFL Removal, basically got rid of a huge module that was slow but accurate but the replacements are as accurate and faster
• NMS-Free, similar to above, introduced in YOLOv10
• ProgLoss+STAL, this is a new technology to help work with small detections and be more accurate with those.
• MuSGD, a new technology used to formulate outputs.
• They claim 43% Faster CPU Inference, which I will talk about later.

Data Provided

Detection

these were steady gains but a tad slower

I also wanted to mention they had released many graphs, however omitted YOLO12, which came in between YOLO11 and YOLO26 and was the SOTA. There was a bit of controversy on this and you can see on the YOLO12 native github.

So I made a graph showing YOLO26 in green, YOLO12 v1 in blue, and YOLO12 turbo in red.

the graph Ultralytics will not show you. it’s ok but larger (YOLO26 in green, YOLO12 v1 in blue, and YOLO12 turbo in red.)

The reality is different: it is not as fast as YOLO12. This was whilst YOLO12 uses FlashAttention which many people pointed as being heavier and slow already.

But the accuracy is higher. This is a tradeoff for those who want to use YOLO12. You will get a faster but less accurate, but sometimes, like how ODVerse33 showed, YOLO12 might be more accurate.

Food for thought.

Segmentation

graph showing YOLO26 segmentation (green) vs YOLO12 SOTA image segmentation (red)

Again it is the same, slower models yet more accuracy.

I stress that people need to test YOLO26 and YOLO12, versus their best models before making a decision on the model they will use and deploy in their projects.

Classification + OBB + Pose + YOLOE

Again, very similar to the above. We are getting slower models but more accuracy.

There was YOLOE released with this, but I need some time to view this and give a comprehensive look (it is pretty different).

Overall Trends

The release of YOLO26 was turbulent. From getting a speculative date to doubts of release to preliminary data that looked subliminal, YOLO26 has had its fair share of controversy with release.

But we are starting to see the turn go on with YOLO models. Last year around this time, I had commented that models were seriously getting slower but more accurate, and groups of researchers who primarily work on mobile applications are being left behind in the past.

But at the same time, technology is manifesting that this models will fit the now more advanced mobile technology.

My main concern however is how Ultralytics handled this. Announcing a model with no date, to randomly announcing release without 24 hours wait, to having models that are slower was rather interesting.

They said “43% faster CPU Inference times”. I don’t know if I am tweaking but I have found nothing to support this claim. They even went ahead and removed YOLO12 from their graphs too which can show nuance and show that YOLO26 is not as good as it seems.

I will probably have an article soon talking about stuff like this, but it is sad to see the YOLO-verse starting to sour. It went from independent workers from Ultralytics working on new versions of YOLO getting fully abandoned and not having their models utilized as much as the Ultralytics natives. Then there was the YOLO13 controversy. And now this.

It is saddening to see on a release date that tactics of hiding non-native YOLO12 from graphs to show huge margins still continue. I am not impressed.

YOLO26 is slower, but more accurate. Do with that as you will.

If you want more articles like this, make sure to follow. Comment your thoughts on YOLO26. Make sure to clap to increase its appearance in search results. I have an email subscription to get an email every week on something really intriguing about health science research fields. That way you don’t forget.

And that’s about it, I will catch you on the flip flop, I’m out, see yah into the new year!!!