At the moment, there’s still no ETA for SAM2 or SAM3 training/fine-tuning support in Ultralytics.
Right now the support is inference-focused: the SAM 2 docs show Training as unsupported, and the SAM 3 docs currently cover predict and track workflows. So for custom SAM training today, the best path is still the official Meta implementation.
If you want, share your exact use case for fine-tuning and we can suggest the closest current workflow.