Ultralytics

Simpler. Smarter. Further.

[Ultralytics @ultralytics]

Ultralytics YOLO Vision 2026 returns September 13 🚀

Join our hybrid global event for the next generation of Ultralytics YOLO, major launches, live demos, breakthrough research, and real-world deployments.

Open vision, built for the real world.

Register now 👉 https://t.co/QRJwNGolbF

[Muhammad Rizwan Munawar @muhammdrizwanmr]

12 additional languages in the @ultralytics Platform! 🌍

Navigate datasets, training, deployments, and settings in your preferred language, including Arabic with right-to-left support.

Select your language from the top menu, and the Platform remembers it.

Explore Ultralytics Platform ➡️ https://t.co/OwUPMDkeLa

#AI #Research #Languages

[Ultralytics @ultralytics]

371 official datasets on Ultralytics Platform, with 53 just added 📚

All 53 allow commercial use, licenses checked at the original release. 29 industrial and electronics, 20 medical and 4 UI datasets: 303K images, 7.4M annotations.

Browse them ➡️ https://t.co/R9NRv1ubPG

[Ultralytics @ultralytics]

How is vision AI transforming real estate? 🏠

REVA is using Ultralytics YOLO to automate license plate redaction across 3,500+ property videos, reducing manual work and making property management more efficient.

Explore the customer story 👉: https://t.co/kRK1SokYcQ

[Ultralytics @ultralytics]

From predictions to precision–recall curves 📈

Learn how mean average precision (mAP) evaluates detection models & why it matters for real-world AI.

➡️ https://t.co/MZfuo9ueCU

[Ultralytics @ultralytics]

Gemini 4 Carbon now tops our auto-annotation leaderboard 🏆

0.584 mAP50 across 36 real-world detection datasets, ahead of GPT-6.1 Terra at 0.560, plus the best mAP50-95 on the board (0.371).

See how it annotates ➡️ https://t.co/0kzrHj3459

[Ultralytics @ultralytics]

New release Ultralytics 8.4.203 | Faster deterministic segmentation loss 🚀

With deterministic=True, semantic segmentation Dice loss now skips a slow CUDA path: 10.9 ms → 1.5 ms forward+backward on an RTX PRO 6000 (batch 8, 19 classes, 256x512).

Explore ➡️ https://t.co/CNbCjKVpCu

#Ultralytics #SemanticSegmentation #PyTorch

[Ultralytics @ultralytics]

Run Ultralytics YOLO26 on MediaTek Dimensity 📱

A new guide shows how to export YOLO26 to LiteRT and run all seven tasks on a Dimensity phone's CPU or GPU. YOLO26n detection takes 25.6 ms per prediction on the GPU of a Xiaomi 17T Pro (640 px, w8a32, mean of 15 runs).

Explore ➡️ https://t.co/3J5VvGovyF

[Ultralytics @ultralytics]

Make home security cameras more than just recording devices. 🏠

With Ultralytics YOLO26, detect people in restricted areas, trigger alerts when the alarm is armed, and monitor activity in real time.

Read more ➡️ https://t.co/81YPASV8nr

#Ultralytics #YOLO26 #ComputerVision

[Ultralytics @ultralytics]

Find similar images with more control! 🔎

Ultralytics Platform now lets you filter similar images by commercial-use licenses or your own images, with license details displayed for each result.

Explore ➡️ https://t.co/jcQ9IHjtsv

[Ultralytics @ultralytics]

Faster depth inference on Snapdragon! 🚀

With refreshed QNN models in the Ultralytics YOLO Flutter plugin v0.7.1, inference on Snapdragon 8 Elite Gen 5 dropped from 26.1 ms to 7.4 ms (~3.5× faster).

QNN export requires Ultralytics v8.4.188+.

🔗 Learn more ➡️ https://t.co/VeE78WFOHQ

#EdgeAI #ComputerVision #YOLO

[Ultralytics @ultralytics]

New tutorial | Missing PPE detection with Ultralytics YOLO26 🦺

Train YOLO26 on Construction-PPE to flag missing helmets, vests, gloves, boots and goggles, labeled only where the body part is actually visible.

Read the guide ➡️ https://t.co/igFKQpmwvV

#Ultralytics #Safety #Construction

[Ultralytics @ultralytics]

OpenAI Images 2.6 is now in Ultralytics Platform! 🎨

Generate training images with GPT-Image-2.6 Flare or Sunburst, added less than 24 hours after OpenAI's release, and turn one dataset image into up to 16 new scenes.

Try it out ➡️ https://t.co/DNwlQWiyQm

[Ultralytics @ultralytics]

Claude Fable 5.2 refused every baggage X-ray in our benchmark, even a scan of an ordinary phone. 🚫

Every open-weights and Chinese model we tested answered all 10.

A closed lab censors. Open weights answer.

See the refusals ➡️ https://t.co/gBkMKFm11z

#Ultralytics #OpenSource #AI

[Muhammad Rizwan Munawar @muhammdrizwanmr]

How do neural networks understand images? Through feature maps. 🧠

Feature maps are intermediate outputs inside convolutional neural networks that highlight important visual patterns in an image. Instead of working directly with raw pixels, the model transforms the image into multiple feature maps that capture edges, textures, shapes, and eventually object-level details.

Early feature maps focus on simple structures, such as lines and corners. Deeper layers combine these into higher-level representations, allowing the model to recognize meaningful objects and regions. This layered understanding is what makes modern computer vision models so effective.

📖 Learn more ➡️ https://t.co/Bzgwh2eYAC

#AI #Research #Mlops

[Ultralytics @ultralytics]

How does AI track objects in specific zones? 🎯

With Ultralytics YOLO26, TrackZone uses trackers like ByteTrack, BoT-SORT, and FastTrack to monitor objects within defined areas in real time.

Learn more ➡️ https://t.co/nbCTOraJGS

#Ultralytics #ObjectTracking #VisionAI

[Ultralytics @ultralytics]

New release Ultralytics 8.4.208 | Fewer duplicate boxes 🚀

YOLO predict and track now default to an NMS IoU of 0.5 instead of 0.7, so the overlap filter drops more near-duplicate boxes of the same class. Validation stays at 0.7; our iOS SDK uses 0.5 too.

Explore ➡️ https://t.co/OIfTwuN5eg

#Ultralytics #YOLO26 #ComputerVision

[Ultralytics @ultralytics]

16 cameras. One NVIDIA DGX Spark. 32 FPS per camera! 🚀

Impressive real-time multi-camera inference with Ultralytics YOLO26s, DeepStream 9.2, and TensorRT FP16, shared by @atasrb.

⚡ 16 simultaneous 1080p streams at 31–32 FPS each!

Check out the setup 👇

➡️ https://t.co/fT4y09K0Px

[ata @atasrb]

One DGX Spark detecting and tracking 16 1080p cameras, every frame, 31-32 FPS per camera.

YOLO26s on DeepStream 9.2, TensorRT FP16.
One Spark tops out near 18 cameras at 30 FPS.

Clips from Pexels. Recipe in the reply.

[Ultralytics @ultralytics]

Ultralytics YOLO26 at @AMD's Embedded AI Working Group in Dublin! 🇮🇪

25+ attendees tried YOLO26 pose estimation in two AMD-led workshops, and our live demos ran all seven YOLO26 tasks on AMD edge hardware.

Read more in the blog ➡️ https://t.co/AlcGf2VbcY

[Muhammad Rizwan Munawar @muhammdrizwanmr]

Cap counting in the manufacturing line using @ultralytics YOLO26 🏭

#Manufacturing #AI #Research

[Ultralytics @ultralytics]

Ultralytics VLM Leaderboard now scores every model on mAP50-95, precision, recall, and F1 📊

Precision is the share of drafted boxes matching labeled objects; recall is the share of labeled objects found. Pick the model that fits how you review.

Compare them ➡️ https://t.co/0a4a1IVlXB

#Ultralytics #VLM #ComputerVision