Ultralytics

What is Ultralytics?

Ultralytics is the company behind Ultralytics YOLO, a family of real-time vision models distributed as an open-source Python package, and Ultralytics Platform, a browser-based computer vision platform for labeling datasets, training YOLO models on cloud GPUs and deploying them to hosted endpoints.

Its terms of service define two services: Ultralytics YOLO Models, computer vision models and related tooling offered under open-source or commercial license terms, and Ultralytics Platform, a cloud-based software-as-a-service environment for dataset management, model training, annotation and deployment. The terms treat these as two distinct product lines, each with its own pricing and licensing structure. The contracting entity is Ultralytics, Inc., a corporation incorporated in Delaware.

In practice, paying for Platform compute and being allowed to ship closed-source software are separate questions.

Core features

YOLO models and the Python package

• Seven tasks, six modes: The package covers object detection, instance segmentation, semantic segmentation, depth estimation, classification, pose estimation and oriented bounding boxes (OBB), and every model can be trained, validated, used for prediction, exported, used for tracking or benchmarked.
• YOLO26 as the current release: YOLO26, which Ultralytics describes as its newest model family, was released in January 2026 and adds native end-to-end inference.
• Optional NMS-free output: End-to-end inference without a separate non-maximum suppression (NMS) pass is available with nms=False, while the default path still uses NMS for accuracy.
• Published benchmarks: Across its five detection sizes, YOLO26 is reported at 40.9-57.5 mAP on COCO with 1.7-11.8 ms latency on a T4 GPU with TensorRT, figures that come from Ultralytics' own paper rather than independent testing.

Ultralytics Platform

• Smart annotation: The annotation editor covers six task types, and smart annotation can use SAM 2.1, SAM 3 or SAM 3.1 (the default), your own fine-tuned YOLO models, or class-prompted models, including paid vision models from providers such as OpenAI, Anthropic and Google when you add your own provider key.
• Cloud GPU training: Cloud training offers 24 GPU types on every plan and 26 on Pro or Enterprise, which add B200 and B300 cards, with real-time metrics and a live cost estimate.
• Remote training: Remote training lets you train on your own hardware while streaming metrics back to the Platform.
• Export: Trained models can be exported to 22 deployment formats, including ONNX, TensorRT, CoreML, LiteRT, Hailo and Ascend.
• Dedicated endpoints: Each dedicated endpoint is a single-tenant service with its own URL and monitoring, and it can be placed in any of 42 deployment regions.
• Agents: Agents workflows connect models, conditions, dataset collection, Slack alerts and webhooks.

The two halves overlap: the package alone can train and export on your own machines, while the Platform hosts the labeling, GPUs and endpoints.

Guide

Run a first prediction with the package

1. Install or update the package with pip install -U ultralytics in a Python 3.8+ environment with PyTorch 1.8 or newer.
2. Run yolo predict model=yolo26n.pt; the pretrained weights download automatically, the model runs on two bundled sample images, and the command prints where the annotated results were saved.
3. Point source= at your own image, video, folder, URL, stream or webcam (source=0) to run the same model on your data.

Train and deploy on Ultralytics Platform

1. Create an account at platform.ultralytics.com and choose a data region (US, EU or AP) during onboarding.
2. Open Annotate in the sidebar, upload your dataset and label it.
3. Create a project and start training on cloud GPUs.
4. Test the model in the browser, then deploy it to a dedicated endpoint.

Ultralytics use cases and examples

Ultralytics markets its models for manufacturing, healthcare, automotive, agriculture, retail and logistics, naming visual inspection, medical imaging, real-time object detection and facial recognition among typical uses. Its published customer stories show what that looks like in production:

• Mining conveyor inspection: In a case study published by Ultralytics, SK Godelius uses YOLO models on robots that walk alongside conveyor belts and flag rollers that are missing, stalled or overheating.
• Reported inspection results: The same study says the system detected more than 3,400 anomalies across its deployment, over 900 of them classified as critical.
• Food assembly robotics: Another vendor case study says Chef Robotics uses YOLO for tray and ingredient detection and segmentation on production lines and reports food giveaway falling from 9.19% to 3.05%.

These figures are vendor-published, not independently measured, and both describe custom detectors inside commercial products, the situation the Enterprise License targets.

Who is it for

• Researchers, students and open-source projects: Ultralytics says AGPL-3.0 suits academic research, coursework, personal projects and open-source applications whose full source code you can publicly release.
• Companies keeping code private: An Enterprise License is required to use Ultralytics YOLO without open-sourcing the entire project, including internal business tools, commercial products, SaaS back ends and embedded devices.
• Individuals, small teams and larger organizations: Ultralytics positions Free for individuals getting started, Pro for professionals and small teams, and Enterprise for organizations that need unlimited training and storage, on-premises deployment, SSO/SAML and dedicated support.

It fits poorly for closed-source teams with no budget for a quoted license.

Platforms

• Web app and data residency: Each Platform account keeps its datasets, models and managed training data in one of three regions chosen at onboarding: US (Iowa), EU (Belgium) or Asia-Pacific (Taiwan).
• Python package: Besides pip, the package can be installed from conda-forge with conda install -c conda-forge ultralytics.
• Edge and mobile targets: Exported models target edge hardware such as NVIDIA Jetson, Raspberry Pi and mobile phones through formats including ONNX, TensorRT, CoreML and TFLite.
• iPhone and iPad: A free Ultralytics YOLO app for iOS 16.0 or later runs detection, segmentation, pose, classification and OBB models on-device through the Apple Neural Engine.
• API: A REST API gives programmatic access to Platform features for automation and CI/CD pipelines.
• Integrations: Integrations connect the Platform to cloud storage, external annotation tools and Slack.
• Cloud storage datasets: Using Google Cloud Storage, Amazon S3 or Azure Blob Storage datasets requires Pro or Enterprise.

Pricing

Software licensing is priced separately from the Free, Pro and Enterprise Platform plans below.

Pro is also sold yearly at $290 per seat, about 17% less than paying monthly, and its $30 per-seat credit is granted every calendar month. The Free signup credit is $5 and rises to $25 once you sign up with, or later verify, a company email address on a non-consumer domain, a bonus granted once per account. Monthly Pro credits are use-it-or-lose-it and expire at each billing-cycle boundary, while credits you buy through top-ups never expire.

Metered compute

• GPU rates: Cloud GPUs are billed per hour from $0.24, Free accounts can reach cards costing up to $4.39 per hour, Pro and Enterprise extend that to $7.39 per hour, and Ultralytics says there is no markup, minimum or commitment.
• Batch annotation: Batch annotation costs $1.00 per 1,000 images, with a minimum of $0.01 per run.
• Endpoints: The default endpoint size of 1 vCPU and 2 GiB is free and scales to zero when idle, while custom sizes keep one warm instance and are charged hourly from readiness until stopped, idle time included.
• Negative balance: If metered usage pushes the balance below zero, a running cloud training job stops and new jobs cannot start until the balance is positive, but deployments keep running.
• Failed jobs: Jobs are charged for elapsed GPU time once a cloud GPU has started, even if they fail or are cancelled, while validation or launch failures before compute starts carry no GPU charge.
• Cancellation: After you cancel Pro, its features stay active until the billing period ends and monthly credits stop at cancellation; once the workspace reverts to Free, all non-owner team members are removed.

Licensing costs

Free and Pro plans operate under AGPL-3.0; the YOLO commercial license for proprietary, closed-source deployment is the Enterprise License, which the Enterprise plan includes. Enterprise plans are provisioned by the Ultralytics team and bundle the Enterprise License with SSO/SAML, on-premise data and compute, and custom SLAs. The Enterprise License is quoted individually, tailored to each organization's size and use case through a request form. Under the Enterprise Software License Terms, the license runs for one year and renews automatically for one-year terms unless either party gives written notice at least 45 days before the term ends.

Ultralytics alternatives

• Roboflow (hosted platform): Roboflow lists a Free Tier at $0 with 10 credits a month and a Core plan at $39 per month, billed monthly.
• Roboflow labeling: Roboflow pitches building, deploying and monitoring vision use cases on one platform, and its AI labeling tools such as Label Assist, Smart Polygon, Box Prompting and Auto Label consume credits.
• Detectron2 (open-source library): Detectron2 is described by its maintainers as Facebook AI Research's next-generation library for state-of-the-art detection and segmentation algorithms.
• Detectron2 licensing: Detectron2 is released under the Apache 2.0 license.

Roboflow is the closer match to Ultralytics Platform; Detectron2 matters mainly when the license, not the tooling, is the reason to switch.

Limitations

Licensing obligations

• AGPL scope: The choice between open-sourcing the whole project under AGPL-3.0 and obtaining an Enterprise License applies even if you train your own model from scratch, skip pretrained weights, use YOLO only internally or for R&D, or deploy it behind a SaaS platform or API.
• Trained models: Models you train with Ultralytics YOLO fall under AGPL-3.0 by default, because the license is described as covering both the training code and the models it produces.
• Downstream projects: In March 2024 the maintainer of Frigate, an MIT-licensed open-source project, wrote that Ultralytics had said Frigate violated AGPL-3.0 because of its MIT license, and that the project could not include any reference to Ultralytics' AGPL-licensed models, code or content.
• License expiry: If an Enterprise License is not renewed, products already sold keep their license rights, while unsold products must stop distribution and internal products must be archived and stop using the software, unless otherwise agreed in writing.

Data, compliance and content rules

• Fixed data region: You cannot change your data region yourself after account creation, so a move has to be requested through support.
• Health data: The services are not designed for protected health information (PHI); without a business associate agreement you must not upload PHI, and only Enterprise customers may be eligible to sign one.
• Biometric consent: The terms prohibit processing likenesses, biometric data or identifiable images of third parties without explicit consent, including facial recognition or gait analysis on people who have not agreed to it.

Operations and consistency

• Cold starts: An endpoint that has scaled to zero can take up to about a minute to answer its first request.
• Request limits: Predictions sent through the Platform API are held to 20 requests per minute, while direct endpoint calls skip that limiter but run on a single instance and reject request bodies over 32 MB.
• Supply-chain incident: Independent security reporting in December 2024 documented that ultralytics versions 8.3.41 and 8.3.42 published to PyPI were compromised to deploy a cryptominer; the founder said they were pulled and replaced with a clean 8.3.43, and a later update cited user reports indicating the attack appeared to continue in versions 8.3.45 and 8.3.46.
• Inconsistent figures: One homepage FAQ answer says models can be deployed to 43 global regions, while the pricing table and documentation say 42.
• Export count: A pricing FAQ answer likewise mentions export to all 21 formats, against 22 on the plan cards and in the documentation.

Privacy and data rights

• Ownership: Uploading data grants Ultralytics a license limited to operating and providing the services, and the terms say your ownership of what you upload is not affected.
• Public datasets: Anything you make public through the Community features becomes available to third parties and may be used by Ultralytics or others for any lawful purpose, including training machine learning models.
• Private content: For customer content uploaded to the Platform, the privacy policy describes Ultralytics as a processor handling datasets, images, videos and annotations to provide services such as training, annotation and model execution.
• Enterprise protection: Enterprise customers get an added contractual commitment that Ultralytics will not use their datasets, model weights or other submissions to train, fine-tune or improve its models or products without prior written consent, though aggregated, anonymized account-activity data is excluded from that restriction.
• Package telemetry: The open-source package sends anonymized usage analytics by default, plus crash reports when the sentry-sdk package is already installed, and setting sync=False stops data from being sent to Google Analytics or Sentry.
• Images: Ultralytics states that the package does not collect, process or view training or inference images.
• Certifications: Every Platform plan is covered by ISO/IEC 27001:2022 certification and a SOC 2 Type I audit, with custom SLAs added on Enterprise.

FAQ

Q1. Is there a free trial, and do I need a credit card?

The Free plan works as the trial: it includes the signup credit described above, and no credit card is required to start.

Q2. Can I change plans in the middle of a billing cycle?

Yes. Upgrades take effect immediately, while downgrades apply at the start of the next billing cycle.

Q3. Can I download the models I train on the Platform?

Yes. Each model page offers its .pt weights for download, and the Export tab holds every completed exported format.

Q4. Is YOLO27 available yet?

Not yet. YOLO27 is still in final R&D, the models are not available, and no launch date has been set.