Anvil

Anvil is a self-hosted, S3-compatible object store designed specifically for AI workloads. It handles large model files like safetensors, gguf, and ONNX by providing model-aware indexing and tensor-level streaming, allowing applications to load only the specific layers or tensors needed for inference rather than downloading entire files. The storage layer uses erasure coding for efficiency and supports multi-region clustering, making it suitable for both local development and production environments powering AI applications.

Unlike general-purpose object stores, Anvil understands AI model formats and enables partial reads, significantly reducing latency and storage usage for large language models and fine-tunes.