As of November 13, 2026, the Hugging Face API listed 27,959 public model repositories that carry the tag or name "uncensored", "abliterated", "heretic" or "decensored". Together they showed 78.8 million recent downloads, 392.5 million all-time downloads and 181,973 likes. Seventy-three percent of them were created in 2026, 55% are GGUF files, and 39% are built on a Qwen base model. Download counts include automated pulls and repeat fetches. They are not a count of users.
Repositories by Term
| Term | Carry the tag | Tag or repository name | Recent downloads | Likes |
|---|---|---|---|---|
| uncensored | 16,671 | 20,487 | 72,976,070 | 163,750 |
| abliterated | 13,313 | 18,132 | 57,002,730 | 113,945 |
| heretic | 7,086 | 9,219 | 17,642,337 | 46,814 |
| decensored | 4,527 | 4,557 | 7,218,742 | 20,152 |
| Any of the four, counted once | 27,959 | 78,759,567 | 181,973 |
The rows overlap, because one repository often carries several terms, so only the last row is a total. Of the 27,959, 13,850 match on both tag and name, 9,138 on name only and 4,971 on tag only. For what the terms mean, see abliterated model and abliterated vs uncensored fine-tune vs jailbreak.
How Concentrated the Downloads Are
- The top 10 repositories account for 32% of recent downloads, the top 100 for 67% and the top 1,000 for 87%.
- The median repository had 119 recent downloads. 13,158 had fewer than 100 and 1,715 had none.
- 12 repositories had more than 1 million recent downloads and 82 had more than 100,000.
- 14,248 repositories, 51%, had no likes.
- The repositories belong to 6,507 accounts. One account that publishes converted copies of other people's models holds 6,640 of them, 24% of the set.
The ranked list is in most downloaded uncensored models, 2026.
By Base Model Family
| Family | Repositories | Share | Recent downloads |
|---|---|---|---|
| Qwen (Alibaba) | 11,015 | 39.4% | 53,326,816 |
| Gemma (Google) | 3,778 | 13.5% | 5,821,153 |
| Llama (Meta) | 2,629 | 9.4% | 1,725,510 |
| Mistral | 821 | 2.9% | 641,860 |
| GLM (Z.ai) | 683 | 2.4% | 950,309 |
| LFM (Liquid AI) | 564 | 2.0% | 1,342,656 |
| DeepSeek | 463 | 1.7% | 2,484,845 |
| Image and video diffusion models | 383 | 1.4% | 1,241,585 |
| gpt-oss (OpenAI) | 358 | 1.3% | 424,670 |
| Phi (Microsoft) | 282 | 1.0% | 86,964 |
| Granite (IBM) | 216 | 0.8% | 75,169 |
| Other or not identifiable | 6,767 | 24.2% | 10,638,030 |
Qwen-based repositories take 68% of recent downloads. The mix has shifted: Llama was the base for 45% of repositories created in 2024 and 4% of those created in 2026, while Qwen went from 18% to 43%.
By File Format
| Format | Repositories | Share | Recent downloads |
|---|---|---|---|
| GGUF (for llama.cpp and apps built on it) | 15,405 | 55.1% | 72,520,304 |
| Safetensors, full weights | 7,249 | 25.9% | 4,225,329 |
| MLX (Apple silicon) | 2,855 | 10.2% | 797,304 |
| EXL2 or EXL3 | 605 | 2.2% | 45,107 |
| Adapter (LoRA or PEFT) | 444 | 1.6% | 244,282 |
| ONNX | 95 | 0.3% | 3,094 |
| Other or not tagged | 1,306 | 4.7% | 924,147 |
GGUF repositories take 92% of recent downloads. Much of the repository count is the same model repackaged: a single edit commonly appears as full weights plus several converted copies under other accounts.
Growth by Creation Date
| Created in | Repositories still public |
|---|---|
| 2023 | 348 |
| 2024 | 2,152 |
| 2025 | 5,136 |
| 2026, to November 13 | 20,323 |
| of which January to March 2026 | 4,209 |
| of which April to June 2026 | 5,970 |
| of which July to September 2026 | 6,224 |
| of which October 1 to November 13, 2026 | 3,920 |
The oldest repository in the set dates from March 2023. The first with "abliterated" in the name were created on May 6, 2024, and the first tagged "heretic" on November 3, 2025. These figures count repositories that still exist, so anything deleted or disabled since is missing and earlier years are understated.
Labels and Access Controls
- 1,475 repositories, 5.3%, carry the platform's "not-for-all-audiences" tag.
- 517 are gated: 391 with automatic approval and 126 with manual approval.
- None is marked disabled, because disabled repositories do not appear in the listing.
What those mechanisms are is covered in Hugging Face content policy and enforcement, 2026.
Methodology
We queried the public Hugging Face models API on November 13, 2026 between 21:33 and 21:35 UTC, paging through every result for eight queries: the tag filter and the name search for each of "uncensored", "abliterated", "heretic" and "decensored". Results were merged by repository ID. A repository is counted if it carries one of the tags, in any capitalisation, or has one of the words in its repository name. 44 results that matched only on the account name were dropped. "Carry the tag" is the API's count for the exact lowercase tag. Recent downloads are the API's downloads field and all-time downloads its downloadsAllTime field. Family is assigned by keyword from the repository's base model tag, or from its name where there is no tag (6,157 repositories have none), so it is a heuristic: derivatives such as WizardLM or Vicuna that do not name their base fall under "Other". Format is assigned from tags, library field and name. The name search is not guaranteed to return every substring match, so the counts are a floor. Shares are our arithmetic. The set includes a small number of image models and may include repositories that use these words in another sense.