Trust & safety

Is LLM Torrents a black market for AI models? No. Here is what we actually do — in plain language, with every claim pointing at a page you can check.

Reviewed: 20 July 2026

In short

LLM Torrents mirrors open-weight, redistributable model files only — the same weights their developers already publish openly. We do not host leaked or proprietary weights, and we never mirror files out of a gated repository. The upstream license travels inside every torrent, we respond to takedowns, and there are no accounts, ratios, invites, or paywalls. We are an open model distribution and preservation layer, not a piracy tracker or an inference API.

What LLM Torrents is

Download the real thing — named quant, verified hash — not a black-box endpoint. LLM Torrents (llmtorrents.com) is an open catalog and a BitTorrent distribution layer for openly-licensed model weights: infrastructure for getting real weights onto your own hardware, plus a public archive that keeps useful open models available over time. It is not an inference API — we do not run the models for you; we help you download and verify the actual files and run them yourself.

What we mirror — and what we never will

We list only openly-licensed weights, and we check the license per artifact before carrying it. Models whose origin is unclear are not listed. That means:

  • No leaked or proprietary weights, ever. If a model's weights are not published openly by their developer under a license that permits redistribution, they are not here.
  • Gated upstream repositories are never re-hosted. Where weights sit behind a click-through or access gate, we do not mirror those files — a gated repository's contents are never re-hosted. Such a model may appear in our catalog as metadata only, with its download state shown honestly (for example “gated upstream”) and a link back to the original — never a magnet or .torrent we don't have the right to serve. If a separately published build of the same model exists openly under a license that permits redistribution, that build can be carried on its own merits and is labeled as exactly what it is.
  • The upstream license travels with the artifact. The upstream license file (and any NOTICE file it ships with) is bundled inside the torrent payload itself, and every listing carries its upstream license and a link to its provenance so you can trace where the weights came from.

Which artifacts we carry for a given family, and why, is spelled out in our Artifact Policy.

What “verified” means here

We are deliberately precise about the word. When a model or torrent page shows a provenance label, it means exactly one of these — and nothing more:

  • “Hash matches official upstream file” — the file we serve is byte-for-byte identical to the official source file (its SHA-256 matches the upstream repository's recorded hash).
  • “Hash matches quantizer-published file” — for a community quantization, the file matches what the quantizer published in their own repository. This is not a claim that we verified the quantization against the base model: we can prove the file matches the quantizer's repo, we cannot prove quantization fidelity to the original weights without reproducible-build evidence.
  • “Base model claimed by quantizer; not independently re-quantized by LLM Torrents” — the lineage is what the quantizer states, not something we re-derived.
  • “LLM Torrents verified this mirrored artifact's hash.” — our own signed manifest records the hash of the exact file we mirror, so you can check the download you received against it.

Every mirrored artifact is covered by an Ed25519-signed manifest recording its SHA-256, and when we pull an artifact for a genuine reason — a superseded quant, a changed upstream license, a bad file — it is delisted and recorded on the revocations page. We do not make quality or “fidelity” claims about a quantization beyond the honest labels above.

Who runs this, and why

Listings are curated and published by a single operator — there are no third-party uploads. That is a deliberate design choice: one accountable uploader means every artifact's provenance is traceable to a known source, rather than an anonymous crowd. The goal is a preservation and verified-distribution layer for open weights: keep useful open models alive, and make the file you download checkable against a hash you can inspect. This is self-funded infrastructure: the site sells nothing, runs no ads, and has no revenue attached to downloads.

We are an independent project. LLM Torrents is not affiliated with, endorsed by, or operated by any model developer (Meta, Alibaba/Qwen, Google, Mistral, DeepSeek, OpenAI, or others). We mirror weights they have chosen to publish openly; we do not speak for them.

What we are not

  • Not a private tracker. There is no ratio requirement, no invites, and no bonus economy. Downloading never requires you to seed, and seeding is entirely voluntary.
  • Not a piracy index. We carry openly-licensed weights only (see above), not arbitrary copyrighted content.
  • Not selling anything. Downloads are free and require no account. There is no paid tier and nothing to buy.
  • Not an inference API. We distribute files; we do not run models as a hosted service.

What happens when someone sends a takedown?

We respond to takedown notices. Our full process — where to send a notice, what to include, and what we do on receipt — is on the takedown / notice-and-action page. In brief: we acknowledge receipt, review promptly (target: within five business days), and if a notice is valid we delist the affected model page, stop serving its .torrent / magnet link and manifest, and cease seeding the material from any infrastructure we operate.

We are honest about one limit of BitTorrent: swarms are decentralized. We can stop our own distribution and stop seeding from our own infrastructure, but we cannot remove data already held by third-party peers we do not control.

Privacy: what other people can see

Torrents here are public swarms. The tracker that coordinates them is anonymous — it requires no account and ties no identity to you — but be clear about what that does and doesn't mean: as with all BitTorrent use, your IP address is visible to the other peers in a swarm and to the tracker, along with the fact that your IP is participating in that specific swarm. This is inherent to the protocol, not a bug or a misconfiguration on our end, and it is outside our control.

If you'd rather your IP not be visible to other peers, seed through a VPN or don't seed that torrent. The seeding safety & setup guide covers exactly what is exposed and how to run through a VPN without silently breaking, and our privacy page covers what the site itself stores (browsing and downloading require no account, and there is no third-party ad/analytics tracker).

Common questions

Why should I trust your hashes?

You don't have to trust ours in isolation. Where a file is an official upstream artifact, its SHA-256 matches the developer's own published file — you can check against the source. Where it's a community quantization, its hash matches the quantizer's published file. On top of that, every artifact we mirror carries an Ed25519-signed manifest recording its SHA-256, and the llmt command-line client will re-verify a file you already have against that manifest without re-downloading it.

Are you affiliated with Meta, Qwen, Google, or the other model makers?

No. LLM Torrents is an independent project and is not affiliated with, endorsed by, or operated by any model developer. We mirror weights those developers have chosen to publish openly under redistributable licenses; we do not speak for them.

Isn't distributing models over torrents just piracy?

BitTorrent is a distribution protocol, not a category of content — it's how many Linux distributions and large open datasets ship. What matters is what is distributed. We carry only weights that are published openly under licenses that permit redistribution, each with a provenance link, and each with its upstream license bundled in the torrent. We do not carry leaked, proprietary, or unclearly-licensed weights, and we act on takedown notices.

What about gated models — can I get those here?

Not the gated files, no. If weights sit behind an access gate on their original home, we do not re-host those files — the model may appear as metadata only, with a link back to the source where you can request access from the people who own that gate. Where a separately published, openly-licensed build of the same model exists (for example a community quantization in an ungated repository whose license permits redistribution), that build may be carried — labeled as what it is, never as the gated original.

Do I need an account, and do I have to seed?

No to both. Browsing and downloading require no account. Seeding is entirely voluntary — there is no ratio requirement and no penalty for downloading without seeding. If you do choose to seed, the seeding safety guide walks through doing it safely.

What happens if this site disappears?

The torrents keep working — swarms don't depend on this website, and anyone holding the files can keep seeding them. The full catalog (metadata, hashes, manifests) is published nightly as a downloadable CC0 dump with a checksum on our open data page, so the record of what was published, under which license, with which hashes, survives independently of us.

Sources & where to check

This page restates our infrastructure and curation practices in plain language; it is not legal advice.