ScaiAtlas

ScaiAtlas is the ScaiLabs model registry — a versioned store for machine-learning models with format variants, namespace-based access control, and peer-to-peer distribution across clusters.

Think of it as the source of truth for "which model, which version, which format, who can pull it, and where the bytes live."

What it gives you#

• Versioned model storage — semantic-versioned releases (1.4.0, 2.0.0-rc1) per model, each an immutable set of files with SHA-256 checksums.
• Format variants — a single version can carry multiple representations (e.g. a safetensors original plus gguf quantizations) without duplicating metadata.
• Namespace access control — every model lives in a namespace with public / tenant / private visibility and per-member roles (owner, admin, member, viewer), backed by ScaiKey identity.
• Resumable transfer — chunked, checksum-verified uploads and range-resumable downloads, proxied through the API so clients never need direct storage credentials.
• P2P distribution — versions can publish a chunk manifest so peers in a cluster seed to each other instead of hammering origin storage.

Three ways to use it#

The mental model in one picture#

1
2

namespace  →  model  →  version  →  variant  →  files
 (poolnoodle)  (noodletron)  (1.0.0)   (gguf-q4_k_m)  (*.gguf, config.json, …)

A namespace owns models; a model has versions; a version has one or more variants (the upload itself is the primary variant); each variant is a set of files with checksums. Read Resource model for the full picture.

Next steps#

• New here? Follow the Quickstart — from install to your first download in a few minutes.
• Building an app? Read the Python SDK reference.
• Wiring another language? Read the REST API reference.
• Curious how it fits together? Read Architecture.