The model delivery layer for AI inference
TokenStorage is the global registry and CDN for AI model weights. Publish or mirror once, then let every authorized inference server pull the exact version it needs from a fast, secure production delivery network.
Private preview for inference platforms, GPU clouds, enterprise AI teams, and edge operators.
acme/support-model:production sha256:90c4d9… Inference starts with a download
Autoscaling compute is fast. Moving 100 GB of weights isn't.
Model artifacts are growing faster than the systems used to distribute them. During a deployment, autoscale event, or failover, many GPU servers can request the same large weights at once. Direct pulls from a public hub or raw object storage can create slow starts, redundant transfers, fragile release workflows, and unnecessary origin load.
Model weights, quantized variants, adapters, and runtime dependencies can be tens or hundreds of gigabytes.
Scale-out events create many concurrent downloads of the same release at exactly the moment capacity is needed.
Inference fleets need private access, immutable versions, traceability, policy enforcement, and predictable global delivery.
Connect an existing source or push artifacts directly to a private TokenStorage repository.
TokenStorage records the release manifest, file hashes, metadata, access policy, and version identity.
Inference servers resolve the requested release and download it through the nearest authorized delivery path.
Your infrastructure receives the exact artifact version it requested, with integrity verification and a complete access trail.
Keep the workflow familiar
TokenStorage is designed to fit the ways inference teams already retrieve artifacts: SDKs, command-line pulls, signed URLs, standard HTTP downloads, and compatible registry APIs. Your servers, disks, and runtime stay exactly where they are.
# Pull a pinned release onto the node
tokenstorage pull acme/support-model:70b-fp8 --dir /models/support-modelfrom tokenstorage import snapshot_download
snapshot_download(
repo_id="acme/support-model",
revision="70b-fp8",
local_dir="/models/support-model"
)# Point an existing workflow at the registry
export HF_ENDPOINT=https://registry.tokenstorage.ai
vllm serve acme/support-model# Any client that speaks HTTP range requests
curl -L --continue-at - \
-H "Authorization: Bearer $TOKENSTORAGE_TOKEN" \
https://registry.tokenstorage.ai/v1/acme/support-model/70b-fp8/model-00001.safetensorsInterface preview. These examples show the intended shape of the client surface, not a shipped API. The CLI, SDK, and compatibility design are being worked out with early-access partners — we would rather match the commands your fleet already runs than ask you to rewrite your deployment path.
You can assemble most of this yourself, and plenty of teams have. The question is whether you want to own it. Everything below is a thing someone on your team ends up maintaining.
Bucket policies, a signed-URL service, and token rotation you build and operate.
Private repositories, organization boundaries, and scoped service tokens.
A naming convention per version, enforced by whoever remembers to enforce it.
Human-readable tags resolving to content digests your deployments can pin.
Opaque objects. Format, quantization, adapters, and licence tracked somewhere else.
Weights, formats, tokenizers, quantizations, adapters, licences, and compatibility.
A CDN tuned for web assets. Range, resume, and concurrency behaviour you verify yourself.
Delivery designed for large sharded artifacts and high-concurrency rollouts.
Cloud IAM plus log plumbing, repeated per account and per region.
One policy surface, with an audit trail for artifact access.
This is not a replacement story. TokenStorage mirrors from existing hubs, buckets, and registries — teams keep using them for discovery, development, and as upstream sources. What changes is the path production pulls from.
TokenStorage is in early development. Every capability below is marked with where it actually stands — nothing here is a shipped feature list.
Deliver model artifacts through a distributed network designed for high-concurrency inference rollouts.
Human-readable tags for teams, cryptographic digests for exact, repeatable deployments.
Resumable, range-based, parallel downloads built for large sharded model files.
Organizations, service tokens, temporary credentials, audit logs, and policy-aware delivery.
Import or mirror models from existing hubs, buckets, registries, and build pipelines without forcing a new authoring workflow.
Weights, formats, tokenizers, quantizations, adapters, licenses, and compatibility — not just generic blobs.
You are being asked to route proprietary model weights through someone else's system. That deserves a straight answer about what is built, what is intended, and what is not decided.
SSO/SAML, SCIM, customer-managed keys, private connectivity, and self-hosted or single-tenant delivery. Which of these land first is a question for early-access partners.
No certification or compliance claims. We hold none, and we will not claim one until it is implemented and independently verifiable. A security architecture page — covering where artifacts are cached, isolation boundaries, cache deletion, geography controls, and the threat model — follows the architecture decisions, not the other way round.
Faster scale-out without hammering your origin.
A consistent, controlled source for large model artifacts across clusters and regions.
Onboard tenant models without a mirror per customer.
Private customer models and faster fleet-wide distribution, without repeated load on a single origin.
One approved release, every environment.
Deliver to multiple clouds, regions, and private environments with centralized policy and auditability.
Large models over links you do not control.
Remote locations on constrained or unreliable connections, with resumable, verifiable delivery.
Ship proprietary weights without a public hub.
Distribute proprietary quantizations, fine-tunes, and LoRA adapters through private repositories.
The same immutable-release machinery can carry agent packages, skills, policy bundles, prompts, tool definitions, and static knowledge packs. Model delivery comes first.
Tell us about your deployment and we will explore whether TokenStorage fits. We are looking for teams with real model-distribution pain and a willingness to shape the product.