osFoundry / Platform
Training
Fine-tune or LoRA-adapt models on your own data — locally, on a workstation, or in the cloud.
What it is
Training in osFoundry lets you adapt a model to your own data instead of relying only on prompts. You can do a full fine-tune or a lighter LoRA adaptation, feeding the model examples from your domain — your support transcripts, your codebase conventions, your writing voice, your labeled records — so it responds the way your work actually needs. It sits alongside the rest of osFoundry, so the model you train is a model you can immediately use in chat, agents, and apps in the same workspace.
Because osFoundry is bring-your-own-key and local-first, training runs where your data and hardware make sense: locally on your own machine, on a dedicated workstation with a capable GPU, or in the cloud when you need more horsepower than your desk can supply. You choose the runtime; osFoundry doesn't force your training data through a reseller or sit a markup on top of the compute. When training runs locally or on your own workstation, your data can stay on your device — nothing has to leave to be turned into an adapted model. LoRA is the pragmatic default for most people: it trains a small set of adapter weights rather than rewriting the whole model, so it's cheaper, faster, and fits on more modest hardware, while a full fine-tune is there when you need the model changed more deeply.
This matters because a prompt can only carry so much, and pasting the same context into every request is expensive and brittle. Training bakes your patterns into the weights once, so the behavior is consistent and doesn't depend on re-explaining yourself each time. Paired with osFoundry's BYOK model — you supply the API keys, there's no per-seat pricing, and your knowledge and data already live in the same workspace — training closes the loop between the data you own and the model you actually run against it.
Key capabilities
- Fine-tune or LoRA-adapt models on your own data — full fine-tuning for deep change, LoRA when you want it cheaper and faster on modest hardware
- Run training locally, on a GPU workstation, or in the cloud — you pick the runtime to match your data and hardware
- Local and workstation training keeps your data on your device — nothing has to leave to produce an adapted model
- Bring your own keys, no per-seat fees — no reseller markup sitting on top of your training or inference compute
- Use the model you trained right away in the same workspace for chat, agents, and apps
Use cases
- ML engineer: LoRA-adapt an open model to your team's codebase conventions on a workstation GPU, then wire it into coding agents in the same workspace
- Domain specialist: fine-tune a model on labeled records and transcripts so it uses your terminology and answers in your house style without long prompts
- Privacy-sensitive team: train locally so proprietary or regulated data never leaves the device, keeping the adapted model entirely under your control
Frequently asked questions
What's the difference between LoRA and full fine-tuning here?
A full fine-tune updates the whole model's weights, changing its behavior more deeply but costing more compute and memory. LoRA trains a small set of adapter weights instead, so it's faster, cheaper, and runs on more modest hardware. osFoundry supports both — LoRA is the practical choice for most tasks, and full fine-tuning is available when you need the model changed more thoroughly.
Does my training data leave my machine?
That depends on where you run the job. When you train locally or on your own workstation, your data can stay on your device — osFoundry is local-first and doesn't require sending it anywhere to produce an adapted model. If you choose to train in the cloud for more compute, the data goes to the runtime you select; you make that trade-off deliberately rather than by default.
Can I use a model I trained right after training it?
Yes. A model you adapt in osFoundry is available in the same workspace, so you can put it to work in chat, agents, and apps without exporting it elsewhere. Because osFoundry is bring-your-own-key with no per-seat fees, there's no reseller markup or extra seat cost sitting between training the model and running against it.