Teaching open-weight AI to see without breaking the bank

After receiving early access to NVIDIA Nemotron 3.5 Lightning, we built a successful proof of concept adding vision capabilities to the new open-weights model. Nemotron 3.5 Lightning is designed for easy customization. In our applied research, we trained only a tiny adapter rather than the full model and aligned existing vision encoders with Nemotron's embedding space. During the experiments, the model progressed from hallucinating answers to accurately recognizing the TNG and NVIDIA logos.

Our multi-modal customization runs on entry-level GPUs, such as a single RTX 6000 Pro, and is very cost-efficient - proving that scaling AI capabilities doesn't require massive infrastructure investments.

A big thank you to NVIDIA for the early access, the highly efficient model, and the support throughout.

Check out our article "Exploring NVIDIA Nemotron 3.5 Lightning: Making it see with little resources" to learn more.