Introducing Boardwalk

Boardwalk is an inference autoscaler. It scales serving capacity with demand while preserving customer control over how each model runs. We are an infrastructure research firm working toward model abundance: making a wide range of models practical to serve at every level of demand.

Model abundance

Specialized models need infrastructure that works for their traffic. A fine-tune used by one team may receive requests only a few times an hour. Another model may need many GPUs at once. Both should be viable to run.

Our research focuses on the systems behind that goal: moving model weights, bringing replicas into service, and allocating capacity efficiently as demand changes.

Autoscaling with control

You choose the model weights, GPU class, context length, KV-cache precision, and how long replicas stay warm. You can reserve capacity for steady traffic and scale beyond it as demand grows. Boardwalk adjusts inference capacity around those choices.

Deploy weights from Hugging Face or call an open model from the catalog through an OpenAI-compatible API. Private deployments are billed for GPU time, with storage billed separately. Catalog models are billed per token.

Read the quickstart, review the serving configuration, or create an account.