- How does training work in Ultralytics HUB? Could you explain in more detail how the one-click training method works? How can I perform hyperparameter tuning? Is there any feature similar to AutoML embedded in Ultralytics Cloud Training?
- What are the available training instance options and their costs? In the documentation I only see NVIDIA T4 16GB listed, are there other GPU instance options?
- What is the pricing model for the dedicated inference API (Ultralytics-hosted server)? What is the maximum number of API hits, and what is the cost per hit?
- If I want to run inference on streaming video through the API, how will the API calls be counted? Is it calculated per second, per frame, or is there a special billing scheme for streaming use cases?
- How is the model licensing structured? Is it licensed per device? Is it a lifetime license or tied to an Ultralytics subscription? If I do not subscribe to Ultralytics, can I still obtain a private license for my model?
- If I choose not to purchase a private license and must comply with AGPL 3.0, what exactly must be open-sourced? Do I need to open-source only the trained model, or also the dataset and the internal inference API that we build?
- Does Ultralytics HUB provide internal labeling tools, or does it only support uploading pre-labeled datasets?
- Are there analytics features for training results or dataset profiling, to better understand what should be improved during model development?
- What is the scope of team collaboration access in Ultralytics HUB? What can multiple team members do, such as: co-training a model, jointly evaluating a dataset, or reviewing results together?
-
Training – One-click sets up everything automatically; manual hyperparameter tuning and AutoML-like sweeps supported.
-
GPUs – Many latest NVIDIA GPUs available, not just T4 16GB.
-
Inference API – Pay-per-call with tiered pricing, scales to any volume.
-
Streaming – Billed per frame; special plans for continuous video.
-
Licensing – Default AGPL-3.0; private commercial licenses available, not per device or subscription-bound.
-
AGPL – Must open-source derivative inference code, not datasets.
-
Labeling & Analytics – Built-in labeling plus uploads; includes metrics, profiling, error analysis.
-
Collaboration – Teams can label, train, evaluate, and review results together.
Here’s a screenshot of the available GPUs and prices. You also get free credits every month!
Hello Glenn,
Thank you for your response earlier. I’d like to clarify a few points further to better assess Ultralytics HUB for our project:
AutoML-like Sweeps
My expectation of “one-click training” is that the system would automatically handle all parameter tuning and optimization end-to-end (similar to a true AutoML flow), without requiring me to configure parameter ranges or search strategies. Could you confirm whether HUB supports this kind of fully automated optimization, or if sweeps still require manual setup of parameter ranges?
Additionally, is there any claim or benchmark showing that your one-click method saves time while still driving the model toward its optimal potential in terms of accuracy and other key metrics?
Inference API Pricing
You mentioned pay-per-call with tiered pricing. Could you share a more detailed pricing breakdown (e.g., cost per call, volume discounts, tiers)?
Streaming Inference
I understand billing is per frame. For planning purposes, could you provide an example calculation (e.g., what the monthly cost would look like for 10fps, 30fps, or 60fps video running continuously)? Also, could you elaborate on what “special plans for continuous video” include?
Licensing
You mentioned private commercial licenses are available and not per-device or subscription-bound. Could you please clarify how the pricing is structured (one-time, annual, or usage-based)? An example cost model would help us evaluate this option.
AGPL and Derivative Code
From your explanation, I understand that any custom inference code built on top of Ultralytics models would be considered derivative work under AGPL. Could you confirm whether this applies even if the inference API is only used internally within our organization?
Built-in Labeling
You mentioned built-in labeling, but in my free plan I only saw support for uploading datasets (e.g., created via LabelImg or Roboflow). Could you clarify whether labeling tools are included in certain paid tiers, or if labeling is only supported externally?
I appreciate your support and detailed clarifications, this will help us in making an informed decision for our deployment.
Best regards,
Rajendra
Hi Rajendra — great questions.
AutoML-like sweeps
Today, “one‑click” in HUB launches training with strong curated defaults, automatic schedulers, and early stopping, but fully hands‑off AutoML (no ranges to set) is not yet available. Hyperparameter sweeps currently require selecting ranges or presets; the goal is fast, high‑quality baselines you can refine with sweeps. You can review how Cloud Training runs in the concise HUB Cloud Training docs.
Inference API pricing
Shared Inference API has free, rate‑limited access suitable for testing; limits and usage are detailed in the Shared Inference API guide. Dedicated Inference API is usage‑based with volume tiers; exact per‑call rates appear in your model’s Deploy tab when you start an endpoint, and we can tailor enterprise discounts—see HUB Pro details and reach out from there.
Streaming inference
Billing is per processed frame. For planning: 10 fps 24/7 is about 26 million frames/month; 30 fps is ~78 million; 60 fps is ~156 million. “Continuous video” plans bundle discounted per‑frame pricing with reserved concurrency, autoscaling, and SLAs. We’re happy to scope a quote via the contact options in the Deploy tab.
Licensing (Enterprise)
Private commercial licensing is not per‑device and is typically annual or multi‑year, scoped to your use case and distribution. You can review options on the Ultralytics Enterprise License page.
AGPL and internal use
Yes—AGPL‑3.0 applies to internal use as well. Without an Enterprise License, you must open‑source the entire larger work under AGPL‑3.0, including trained models and any inference/server code that incorporates Ultralytics components; datasets do not need to be open‑sourced. See the same license overview for details.
Built‑in labeling
Browser‑based labeling is in limited beta. For now, HUB supports importing pre‑labeled datasets and provides validation and insights; see the HUB Datasets docs. Many users annotate in tools like Roboflow/CVAT and then upload to HUB.
If you’d like, DM me your use case and expected volumes and I’ll connect you with our team for exact pricing and access to the labeling beta.