Home / Features / Train and fine-tune / Train custom models

Train custom AI models on osFoundry

osFoundry trains custom adapters and fine-tunes for any of 60+ open-weight base models. Use your knowledge bases, uploaded datasets, or any of 250K public datasets indexed from HuggingFace. Training jobs run on local GPUs, in osFoundry cloud, or against your own infrastructure — your choice per job.

Quick answer

LoRA + QLoRA training on 60+ base models.
Three training paths: local GPU, osFoundry cloud, BYO infrastructure.
Train on KBs, uploaded datasets, or public datasets.
Trained models are workspace-routable instantly.

Frequently asked questions

Can I train on a local GPU?

Yes. Local training runs on Apple Silicon and on NVIDIA GPUs. Note that inference and training have different requirements: 4-bit QLoRA needs an NVIDIA GPU (via bitsandbytes), while LoRA and full fine-tunes also run on Apple Silicon and CPU.

Does osFoundry support reinforcement learning?

RLHF/DPO is on the roadmap. Today, supervised LoRA + QLoRA are the supported flows.

Related features