Artifact Policy

What we carry for each model family, and why — plus how to ask for more.

Last updated: 10 July 2026

Download the real thing

Download the real thing — named quant, verified hash — not a black-box endpoint. Every artifact we distribute is a specific, named quantization or source weights file, with a verified hash and a torrent you can inspect. We are an open model distribution and public model archive — infrastructure for getting real weights onto your own hardware, not an inference API.

The four-rung artifact ladder

To keep seed density high and the catalog honest, every carried model family publishes a bounded, declared set of artifacts rather than every quant variant that exists upstream. For each family we carry, by default, up to four rungs:

  1. Source weights — BF16/F16 or the original safetensors, carried only where the upstream license permits redistribution.
  2. One default consumer quant — a Q4_K_M-class artifact (including dynamic-quant equivalents such as UD-Q4_K_M), the balanced size/quality pick most people should reach for first.
  3. One high-value-per-byte low-bit artifact — an IQ4_XS-class or strong dynamic low-bit quant, for tighter hardware.
  4. One upper-tier artifact — a Q8_0-class quant, for readers who want closer-to-source quality and have the room for it.

Everything else — other quant levels, alternate formats, imatrix variants — is opt-in and demand-driven: we carry it once there's a real request for it, not by default. Seed density on a small, well-chosen set of artifacts beats a wide catalog that seeds thinly across dozens of near-duplicate quants.

The giants preservation lane

Very large model families (roughly 235B parameters and up) are carried as a preservation/showcase lane, not as the default journey through the catalog. We aim to keep at least one verifiable artifact of a giant model available, but we do not promise the same four-rung ladder, and we do not steer readers toward a giant as their "recommended" pick unless their own hardware profile earns it.

Requesting more quants

If the ladder doesn't cover the artifact you need — a different quant level, an alternate format, or a family we don't carry yet — use the request board to ask for it. Demand there is what drives what we add beyond the default ladder.

Why this policy exists

A bounded, declared ladder keeps the catalog legible: you know what to expect for any family we carry, and you know it was chosen on purpose rather than mirrored wholesale. It also keeps seeders concentrated on a small set of artifacts instead of spread thin across every quant variant that exists upstream, which is what keeps downloads fast.

This is an infrastructure and curation policy, not legal advice. See our Terms of Service for how model licenses apply to what you download.