Player with Ball Detection
Detect which player is currently holding the basketball in NBA game footage. The base model fires dozens of false-positive boxes. After reinforcement learning fine-tuning, the model locks onto the correct player with an F1 of 0.79, compared to 0.53 from GPT-5.4.
What's a finetune?
A finetune is a custom-trained version of Moondream, optimized for your specific task using your own data. It starts from the base model and learns to perform better on exactly the kind of images and questions you care about.
Base Moondream 3 Preview
Fine-tuned Moondream 3 Preview
Results
See it in action
Side-by-side comparisons on real game footage. Base model on the left, fine-tuned on the right.
Base Moondream 3 Preview
Fine-tuned Moondream 3 Preview
Getting Started
What is fine-tuning?
Fine-tuning takes a general-purpose vision model and trains it on your specific task. You provide examples of what you want the model to recognize, and the model learns to do that one thing very well. The result is a custom model that performs far better on your task than the base model does out of the box.
Who is this for?
Anyone building a product or system that needs to understand images. You do not need a machine learning background. If you can collect example images and describe what you want the model to see, you can fine-tune Lens.
For teams that want hands-on help, we also offer a white glove fine-tuning service.
How it works
Prepare your data.
Collect images that represent your task. Label them with the outputs you want. For object detection, that means drawing bounding boxes. For classification, that means assigning categories.
Train with the API.
Send your data to the Moondream Lens API. Choose SFT to teach the model from labeled examples, or RL to optimize outputs against a scoring function. Training runs on our infrastructure. There is no hardware to manage.
Deploy your model.
Your fine-tuned model is ready to use through the Moondream API or through Photon, our self-hosted inference engine. Run it in the cloud or on your own hardware.
Code
See the code
These examples show the minimum Python you need to run this workflow with SFT or RL.
import moondream as md
lens = md.Lens()
# Train on labeled examples for this task.
lens.train_step(
training_data=[
{
"image": image,
"prompt": "player with the ball",
"output": "Bounding box around the ball handler",
}
for image in training_images
]
)Why Moondream
The fastest path from idea to production
The point is not to learn another platform. The point is to get a custom model into production with as little friction as possible.
Fully hosted.
API-only.
Fine-tuning is a handful of API calls. There is no UI to learn, no platform to onboard to, and no proprietary format to adopt. It fits into the workflow you already have.
Pay as you go.
You pay for the compute you use. Fine-tuning starts at a few dollars. Every account gets $5 of free credits each month, so you can run your first fine-tune at no cost.
Built by the model team.
Moondream's fine-tuning system is built by the same team that designed the model architecture. The training pipeline is optimized specifically for Moondream.
White Glove
Need help? We'll build it for you.
Not every team has the bandwidth to run a fine-tune in-house. If you want help, our team can handle the process end to end.
Our team works with you to define the task, prepare the data, run the training, and validate the results. When we are done, we hand off everything: the fine-tuned model, the training data, the evaluation benchmarks, and documentation on how to maintain and improve the model over time.
You own the model. You own the data. We just get you there faster.
Task definition and benchmark design.
Data review, preparation, and labeling guidance.
Training, evaluation, and handoff documentation.
A model your team can run through the API or Photon.
FAQ
Frequently asked questions
Concise answers for teams evaluating Moondream for production fine-tuning.
Ready to take Moondream to production?
Every Moondream account includes $5 of free credits per month. No credit card required.
Start fine-tuning
Start with the docs and run your first experiment in a few API calls.
Start fine-tuningTalk to our team
Tell us what you are building. We can help with data, training, evaluation, and deployment.