PARTLY

Partly true: Pirate Face is a real torrent index of open-weight Hugging Face models with a new 'uncensored' filter, but it is not 'every jailbroken LLM', nothing is 'cracked', the Pliny tie is a repost, and 'comment CRACK' gates a public website

Facebook · Oct 2026

The claim'If you want every single jailbroken LLM, well, there's a hub for all of them.' 'Pliny, the number one jailbreaker, is associated with them.' 'You have all LLMs that are local and cracked. There's now a direct download and a centralized hub for all of them.' On screen: 'ALL LLM CRACKED (FREEDOM AT LAST)' over an X post reading 'Abliterated, Heretic, and Obliteratus models are now fully discoverable on Pirate Face.' 'Comment crack and I'll send it over.'

There is a real site under this. Pirate Face (pirateface.co) indexes Hugging Face models as BitTorrent magnets, with each file checked against Hugging Face's own SHA-256 hash and a web seed pointing back to the original. It launched its own tracker on September 20 and says 669,000 models are eligible. It recently added an 'Uncensored' filter with three tags, abliterated, heretic and obliteratus, which are methods for editing an open model's weights so it stops refusing. That is the news item. The reel's wording is not. 'Every single jailbroken LLM' and 'all LLM cracked' are wrong on both ends: the site says it mirrors only Apache-2.0 and MIT licensed models plus a couple of named exceptions, so closed models such as GPT, Claude and Gemini are not there and cannot be, and open-weight models are not 'cracked' because their makers published the weights. 'Pliny is associated with them' rests on a screenshot of Pliny the Liberator reposting the announcement. One of the three tags is named for OBLITERATUS, his open-source toolkit, but the site names no operators and we found no stated affiliation. 'Direct download' means torrents, which expose your IP address to other peers, as the site's own text says. A hash match proves the file is the one on Hugging Face, not that an anonymous uploader's modified model is safe. And 'comment CRACK' is a DM gate on a public URL.

What holds up

Reel read frame by frame: an X post by Pliny the Liberator ('I LOVE THE SMELL OF FREEDOM IN THE MORNING') quoting Pirate Face's post 'Abliterated, Heretic, and Obliteratus models are now fully discoverable on Pirate Face'; the site's filter panel ('Uncensored models: abliterated, heretic, obliteratus'; 'Made in: United States 1, China 17'); a model list with Magnet buttons (Qwen3.8-27B variants); on-screen 'Comment crack I'll send'.
pirateface.co fetched Oct 4: 'Turn AI into torrents that live forever', '669k+ eligible models', Apache-2.0 and MIT only with listed exceptions, a takedown page, an 'Uncensored' facet, and the text 'BitTorrent can expose my IP address to peers'; no operators named and no mention of Pliny.
Trade coverage (Runtime Wire, Sep 20, 2026): first-party tracker and DHT discovery launched that day; 4,891 witnessed records at launch against the 669,000 eligible; a listing 'does not establish that the model is safe, accurate or lawful to redistribute.'
github.com/elder-plinius/OBLITERATUS read: an open-source abliteration toolkit (8.6k stars, AGPL-3.0) under Pliny's account; that explains the tag name and the repost, not an ownership link.
The poster's sell is a comment-for-DM hook for a site anyone can open; x-ref #422 (same page, same hook, the day before).

What doesn’t

'Every jailbroken LLM' is false: open-licensed, open-weight models only; no closed model is or can be on it.
'Cracked' misdescribes abliteration, which edits published weights; nothing is broken into.
The Pliny 'association' is a repost plus a tag named after his tool; no affiliation is stated anywhere we could read.
Torrenting exposes your IP to peers, and a matching hash does not make an anonymous modified model safe to run.

The catch

If you want open models that survive a takedown, Pirate Face is a real and fairly careful mirror, and you do not need to comment anything to use it. If you want a model without refusals, understand what you are downloading: someone else's edit of the weights, with no one vouching for how it behaves.

How to actually do it

Open the site directly and read its license and takedown pages before you fetch anything; skip the DM.
Prefer the original maker's upload on Hugging Face when it exists, and compare the SHA-256 the torrent lists with the one on the model card.
Check the hardware line first: a 27B model quantized to 4 bits needs roughly 16 to 20 GB of memory, and bigger ones need a workstation.
If you need fewer refusals for legitimate work such as security research or fiction, pick a documented abliteration from a known uploader, run it offline, and keep it away from anything connected to customer data.

Pirate Face and its new uncensored filter are real, and the mirror is limited to openly licensed models. The reel inflates that into 'every LLM cracked', leans on a repost to claim a famous backer, and hides a public URL behind a comment hook.

Confidence: High
Posted by: an AI-news reels creator, ~6k followers; reel sent to Buddy Sat Oct 3, 2026, ~4:26 PM

See the original claim →

We test hype for free. We build the real thing for a living.

Thirty minutes, no pitch — and you'll leave with something useful either way.

Book a call with Todd or start with the free Business Checkup →

The Verdict Weekly

Three verdicts every Friday. Free forever, unsubscribe anytime, no spam — that would be ironic.