Dreamscale Labs is releasing Training on Dreamscale for robotics teams that need to adapt vision-language-action models to their own data. A user picks a base model, with π0.5 supported first, imports a dataset from Hugging Face or a local source, and starts training. The product analyzes the model and the dataset, then auto-configures a recommended fine-tuning recipe, while still allowing changes to settings such as step count. The release lands while the field has not settled on a standard fine-tuning method for these models, so teams still have to choose a recipe themselves. The same platform already runs these models from cloud GPUs. A robot sends camera frames and joint positions and receives actions, which is how Dreamscale serves policies such as DreamZero, MolmoAct2, and π0.5 without a large onboard GPU. The company reports 40 to 50 percent lower tail latency than existing cloud providers, and 320 ms for DreamZero 14B versus 350 ms for NVIDIA's implementation at half the GPU count. For latency-sensitive applications, Real-Time Chunking can be turned on with a toggle. Dreamscale is a San Francisco company in Y Combinator's Fall 2026 batch, founded by Eric Ren, Kevin Lin, Antoine Nguyen, and Chris Yoo. They have focused on keeping large robot models inside motion deadlines once inference moves off the robot. Training gives teams with their own data a checkpoint tuned for a specific robot and task, and a cloud runtime that already accepts custom checkpoints.