Catalog 001 — Model Supply
Curated open models — frontier LLMs, image and video generation, speech, embeddings — with every runtime needed to use them. Deduplicated, license-audited, checksummed. Works on a machine that has never touched the internet.
One precision per model. Newest generation only. No forty near-identical quants of the same checkpoint, no abandoned repos. Every file checksummed, the whole manifest published.
Apache, MIT, and open community licenses only, with license text shipped beside the weights. Anything forbidding redistribution is left off, however good it is.
Runtimes and Python itself are on the drive for every platform. First token in minutes on an air-gapped machine — not after a 90 GB download you can't make.
Both editions ship the same software stack and reference library. The 10 TB edition adds the frontier models exactly as their labs published them — original release weights, in the release precision, at full size.
The complete manifest. BOTH ships on every drive; 10 TB is exclusive to the larger edition. Every size below comes from the build manifest — an exact pinned revision and SHA-256 for every file.
$25/year — optional, never required
Open models move fast. A refresh license gives you a signed feed of everything that changed since your drive was built, plus a one-command tool that pulls only the differences — new models, new generations, updated runtimes — onto your drive from any machine with a connection.
Yes — that's what the curation is for. Everything shipped is under a license permitting redistribution: Apache-2.0 and MIT for most, plus community licenses (Gemma, Stability, Qwen) where redistribution is allowed provided the license text travels with the weights, which it does. Non-commercial and no-redistribution models are excluded regardless of quality.
A snapshot is the point — it can't be revoked, rug-pulled, or silently edited. But models do move, so there's an optional $25/yr refresh license that delta-syncs new material onto your drive. Optional, and the drive never stops working without it.
No. We don't take drives back for reflashing — shipping and handling cost more than the service is worth, and we'd rather say so than offer something we'd do badly. Updates happen over the network via the refresh license, or by hand using the guide on the drive.
Everything on the 1 TB edition runs on 8–32 GB of RAM or VRAM, and the manifest spans that whole range — small models for laptops up through 100B+ mixture-of-experts for workstations. The 10 TB edition's original-release frontier weights target multi-GPU rigs, researchers, and preservation; it also contains every quantized model from the 1 TB edition, so it runs fine on modest hardware too.
NVIDIA's licenses don't permit us to redistribute the driver or toolkit. The drive ships CPU-ready runtimes that work immediately, plus a short offline guide for adding GPU support on a machine that doesn't already have it.
1 TB edition: NVMe SSD in a USB-C enclosure, fast enough to run models directly off the drive. 10 TB edition: 3.5" HDD with a USB-C dock. Standard hardware, plain exFAT, no proprietary format and no DRM. If we vanish, the drive keeps working.
Air-gapped and off-grid environments. Researchers who need provenance-clean weights. Institutions that can't send data to a cloud. Anyone who thinks a physical copy of the open-model era is worth owning before licenses change and repos disappear.
Yes, and we'll show you how — the guide is on the drive. What you're paying for is the curation, the license audit, the verification that each model runs, and not spending a week of bandwidth assembling it. If you have the time and the pipe, do it yourself with our blessing.