surface
Train a model without owning the GPU.
Upload a curated image or video dataset, choose an open-weight base model, train a private LoRA adapter, and evaluate the result from a mobile-first control surface.
lab / training system / wip
Model Studio
A mobile-first control plane for cost-efficient, resumable training and evaluation of private LoRA adapters across compatible compute environments.
under the surface ↓runtime deployment in progress
datasetcheckpoint relayGPU sessionprivate LoRA
The compute session can end. The training state persists, restores on compatible compute, and continues.
system
Cost-efficient compute. Persistent state.
Modal provides the primary GPU training environment, while Hugging Face holds private datasets, model output, and checkpoint state. Checkpointed training can survive an interrupted or bounded compute session and continue without restarting from zero.
status
Public context. Runtime still connecting.
The training system is being connected to its external compute and deployment surfaces. This page intentionally shows the work before the live training runtime is opened publicly.
underneath
The GPU is temporary. The experiment is not.
Model Studio is part of the Warp the Surface lab. The training application keeps its own compute, authentication, checkpoint, and deployment boundaries, allowing the public experience to stay focused while the underlying system evolves independently.
Current system: Gradio + Hugging Face + Modal + AI Toolkit. The first production validation is a bounded FLUX.2 [klein] LoRA run that deliberately stops, restores its saved state on a fresh worker, resumes, publishes the adapter, and generates from the result.
wip
Visible now. Interactive next.
Model Studio is under active development. Direct training access will be added after Modal authentication, compute budget controls, checkpoint relay, and end-to-end inference have been verified in production.