AI @ultralytics, PhD @metu_odtu. author of YOLO26/YOLO27, UltraViT and SAHI. 20M+ open-source downloads, 5 patents, 2200+ citations.
Joined February 2020
- Tweets1.8K
- Following121
- Followers1.4K
- Likes4.2K
Pinned Tweet
I have been working on this for a very long time, finally it's live on Arxiv!
Introducing the long-awaited @Ultralytics YOLO26 paper: state-of-the-art real-time object detection, instance segmentation, image classification, keypoint detection, oriented bounding boxes detection and open-vocabulary detection models that run on any edge device 🔥
Meet me at @CVPR booth #513 if you have questions about the paper!
It was a lot of fun collaborating with my coauthors @arttestt and others.
Of course, it wouldn't be possible without the ultralytics package maintainers like @onuralpszr and rest of the team.
fatih c. akyon retweeted
🚀 DeepSeek-V4.1-Flash-1017 is here!
Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API.
🛠️ Optimized for @DeepSeekHarness, Tencent CodeBuddy, WorkBuddy, and more. Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model.
1/4
fatih c. akyon retweeted
@Google's Gemini 4 Argon is now available in Ultralytics Platform! 🚀
In our latest internal benchmark of 26 AI models across 36 real-world object detection datasets, Gemini 4 Argon ranked #1 in auto-annotation accuracy with 0.565 mAP50, outperforming Claude Opus 5.5 and GPT-6 Astra.
From medical imaging to retail, traffic & industrial inspection, developers can use Gemini 4 Argon to generate bounding boxes, review predictions & annotate datasets at scale, reducing manual labeling effort.
Read more ➡️ ultralytics.com/blog/gemini-…
#Gemini4 #Ultralytics #AI
fatih c. akyon retweeted
I'm using and contributing @fedora for over 15 years. Always give back.💙
I got a new Dell XPS 16 and saw how @OmarchyLinux, Intel and Dell brought Panther Lake to Linux so beautifully, I wanted to bring that care to @fedora with @kdecommunity Plasma.
github.com/onuralpszr/xps-fe…
fatih c. akyon retweeted
Introducing EmbeddingGemma 2! 🚀
Our lightweight, multimodal embedding model maps text, code, images, video, and audio into a single, unified embedding space. Optimized for on-device use cases, it features:
- 740M parameter form factor with modular encoders
- Flexible dimension sizes (768dim-128dim) via Matryoshka Representation Learning (MRL)
- 8K context window (4x larger than text-only EmbeddingGemma)
- A commercially permissive Apache 2.0 license
fatih c. akyon retweeted
Your project architecture, now on an interactive canvas.
See how your services connect.
Click a node to inspect it, restart it, or stop it - directly from the canvas.
Open your project → Topology.
Available in Openship 0.8.0
It’s been a few weeks since our last update, but we’ve been working on the next big step for OpenShip:
You could already deploy your product with OpenShip.
Now, you can scale it across servers as it grows.
OpenShip now handles:
> Automated cluster setup
> Private networking
> Application replicas (PostgreSQL & Redis)
> Shared storage
> Backups & recovery
- Your coding agent can now be your DevOps engineer.
Through MCP, Codex and Claude Code can set up clusters, deploy apps, scale instances, inspect logs, and manage backups.
- The new Services Canvas brings your infrastructure into one view.
See your apps, databases, running instances, and routes together. Open a service, inspect it, and make changes right there.
And meet Ship SDK...
We refactored the core so deployment can become part of your own code.
Build workflows in JavaScript or TypeScript, control an existing OpenShip instance, or embed the engine directly into your product.
- And finally, OpenShip Cloud is also live for beta testing.
Our goal is a smooth deployment experience with managed environments and fair pricing, for anyone who wants to focus on their product without maintaining servers.
This release also improves the everyday work: environment management, Compose migration, routing, backups, rollbacks, and deployment logs.
OpenShip remains open source and self-hostable.
Deploy with OpenShip. Scale with OpenShip.
fatih c. akyon retweeted
Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions.
CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks.
With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks.
We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed.
Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size.
📄 Blog: contrastive-lm.notion.site
💻 Code: github.com/Contrastive-LM/CL…
🗣️ Discord: discord.gg/5dAQEDJBs
🤗 Data & Models: huggingface.co/Contrastive-L…
More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵
fatih c. akyon retweeted
Nemotron-3-Diarization: our latest SOTA diarization model, as ranked on Voice Arena Diarization.
🗣️up to 8 speakers
🌎21 languages
🔥15000 RTFx
⚡️offline and online
🕕variable latency down to 0.32s
💰commercial and non-commercial use
huggingface.co/blog/nvidia/n…
fatih c. akyon retweeted
Deploying DiffusionGemma-Jev (djev) just got a lot easier. You can now spin up a Jev API-compatible endpoint on Google Cloud Run using a single command.
Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec.
It's a straightforward way to experiment without needing your own GPU. Runs at roughly $3/hr and drops to $0 when idle.
Get the code and instructions here: github.com/taeold/djev-run
fatih c. akyon retweeted
Today we’re releasing Adam and Eve, the most human-like AI voices ever built.
Adam ranks #1 among AI models on @DesignArena’s AudioRealismBench.
We believe we’ve crossed the uncanny valley. API access on our Website!
Listen: freyavoice.ai
Leaderboard: designarena.ai/leaderboard/a…
Special thanks to @alpsencerozturk, @ahmeterdempmk, and the entire Freya team for making this possible.
We’re just getting started. Join us!
fatih c. akyon retweeted
London, see you this Wednesday! 🚀
Join us on September 23 for the Ultralytics London Meetup, bringing together the vision AI community for an evening of insights, live demos, food, drinks, and networking.
Hear from Ultralytics’ Founder and CEO to discover what’s coming next in vision AI, get insights from NVIDIA and Arm on real-world vision AI applications, and connect with the people building it.
Register ➡️ bit.ly/4rlbzD8
fatih c. akyon retweeted
pleased to share the entire arXiv site as a dataset on HuggingFace
huggingface.co/datasets/sece…
3,148,796 papers, every version, in LaTeX, PDFs, PostScript, HTML, 16 TB in total
fatih c. akyon retweeted
Turns out everyone's been initializing LoRA weights suboptimally.
We find that by using the singular value decomposition (SVD) of the weight matrix to determine both the frozen weights and LoRA initialization, we can achieve faster convergence in RL training.
See the full post below on our extension of PiSSA
fatih c. akyon retweeted
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
2 days ago, Glenn mentioned a new image encoder called UltraViT during the YOLO27 announcement. It's a better and more efficient encoder than DINOv3/FastViT.
I've been designing this architecture since we published YOLO26! After a big collab with Esat and Jing, we made it the heart of YOLO27.
If you want to learn more about our upcoming UltraViT image encoder, keep an eye on the docs page: docs.ultralytics.com/models/…
Ultralytics YOLO27 is coming soon. 🚀
Announced yesterday at YOLO Vision 2026, YOLO27 models are the most capable YOLO models we've built to date.
✅ Simpler model lineup: Four models, four clear answers. n and s for edge devices, m for demanding detection tasks that still need real-time speed, l for high-precision and safety-critical systems. No x model to evaluate as yolo27l already delivers flagship accuracy.
✅ Hybrid architecture: The best architecture for each model size, rather than one design stretched across the range. Small models stay fast on limited hardware; large models achieve higher accuracy.
✅ Breakthrough accuracy: YOLO27l is the first Ultralytics model past 60 mAP on COCO. Higher accuracy across the range, on the same hardware and at the same speed.
✅ Seven tasks: Object detection, instance and semantic segmentation, depth estimation, image classification, pose estimation, and oriented object detection. 28 models across four sizes, all sharing the same interface.
We're also launching YOLOE-27 (detects from prompts) and YOLO27 Enterprise models (pre-trained for your domain).
Available later this year.
Join the waitlist ➡️ ultralytics.com/yolo/yolo27?…
fatih c. akyon retweeted
Introducing ToolGrad: an efficient framework for generating tool-use datasets. By generating ground-truth tool-use chains before prompts, it achieves a nearly 100% pass rate and improved LLM tool-use performance. More: goo.gle/4xNKTNH
ALT While prior art generates tool-use datasets by searching solutions of user queries with low pass rate, ToolGrad generates successful tool-use chains before generating prompts, yielding high pass rate.
