osFoundry / Platform
Hot-swap LoRA
Swap LoRA adapters on a running base model in under 50ms. Different personas, same weights.
What it is
Hot-swap LoRA lets you switch between fine-tuned model personalities in under 50 milliseconds without reloading the base model. A LoRA (Low-Rank Adaptation) adapter is a small set of weight deltas — typically a few megabytes — that steers a much larger base model toward a specific tone, task, or domain. Because the adapter is tiny and the base weights are shared, you can hold one base model in memory and flip between dozens of adapters as fast as changing a setting. Different personas, same weights.\n\nInside osFoundry this runs entirely on your device. osFoundry is a local-first desktop AI workspace, and its local runtime keeps the base model resident in memory while adapters attach and detach on top of it. Loading a fresh multi-gigabyte model from disk takes seconds and a full GPU warm-up; swapping a LoRA adapter onto weights that are already loaded skips all of that, which is where the sub-50ms figure comes from. You point a chat, an agent, or a pipeline role at an adapter, and the next message runs through that persona — a customer-support voice, a code reviewer, a legal-summary style — without tearing down and rebuilding the model. Switch again and you are back to the base or onto another adapter just as quickly.\n\nThis matters because it turns one downloaded base model into a whole shelf of specialists at almost no cost. You are not paying for, downloading, or GPU-loading a separate model per persona, and because osFoundry is bring-your-own-key and local-first, there are no per-seat fees and no reseller markup — you supply the base model and the adapters, and inference stays on your machine. It is a practical fit for multi-agent setups where each agent needs its own behavior, and for privacy-sensitive work where the data driving each persona should never leave the device.
Key capabilities
- Swap LoRA adapters on an already-loaded base model in under 50ms, with no reload of the multi-gigabyte base weights
- Keep one base model resident in memory and layer many low-rank adapters on top, so switching persona costs a pointer swap, not a cold start
- Run every persona on-device in osFoundry's local runtime, so your prompts, adapters, and outputs never leave your machine
- Assign a different fine-tuned adapter to each agent, chat, or pipeline role and switch between them mid-session
- Bring your own base models and your own adapters with no per-seat fees and no reseller markup on inference
Use cases
- Solo builder: keeps one 7B base model loaded and swaps between a coding adapter, a writing adapter, and a summarizer adapter across tabs without waiting for a model to reload each time
- Multi-agent orchestrator: assigns a distinct fine-tuned LoRA to each agent in a pipeline so a researcher, a critic, and an editor each run in their own voice off the same shared base weights
- Privacy-conscious team: runs domain-specific personas fully on-device, so the sensitive prompts and outputs that shape each persona never leave the local machine
Frequently asked questions
How is hot-swapping a LoRA adapter faster than loading a different model?
Loading a different model reads multi-gigabyte weights from disk and warms up the GPU, which takes seconds. A LoRA adapter is only a few megabytes of low-rank deltas that attach on top of base weights already resident in memory, so the swap is a small in-memory operation rather than a cold start — under 50ms in osFoundry's local runtime.
Do I need a separate base model for each persona?
No. That is the point of the feature. One base model stays loaded, and each persona is just a different adapter layered on the same weights. You download and hold the base once, then keep a shelf of small adapters and switch between them, instead of storing and GPU-loading a full model per persona.
Does the adapter swapping happen on my machine or in the cloud?
On your machine. osFoundry is a local-first desktop workspace and runs local models through its on-device runtime, so the base model, the adapters, and the data flowing through each persona stay on your device. It is bring-your-own-key with no per-seat fees and no markup on inference.