62×
growth since 2024
From 342 catalog models created by Jan 2024 to 21,091 today, by Hugging Face creation date
21,093 models
5,547 authors
3,664 base models
+53 new · 24 h
+490 new · 7 days
+2,016 new · 30 days
FLAGSHIP INDEX
19.1%
Disenchantment
Index
Share of the market that has flowed into abliterated versions. 124.4x growth in 24 months.
See Indices tab
NOW snapshot of what the market looks like right now

This month's leaders

Three views of what dominated the last 30 days: the single most-downloaded new release, the most-productive author, and the family that grew the fastest. All updated continuously.
EXPLORE the full timeline →

All-time downloads

The ten most-downloaded abliterations in the catalog by lifetime downloads, since each first appeared: the Hugging Face all-time count from the hub-stats snapshot, or the last-30-days counter when that is higher (a lifetime count can never be lower). This is cumulative popularity, not monthly velocity.

Current momentum

How many new abliterations of each family appeared in the last 30, 60, and 90 days. The trend arrow compares the most recent 30-day window against the 30 days before it: ↑ means the pace accelerated, ↓ means it slowed, → means about the same. Shows which families the community is actively working on right now.

Community reaction time

How fast the community responds when a new base model is released. The first column averages, over each family's original base models, the Freedom Velocity days: from a base model's release on Hugging Face to the first abliterated or uncensored catalog model built on it. Also: how many abliterations that family accumulated in its first 30 days, and how many are being made per month right now. Faster reaction and higher current rate both mean the family is a popular target.
MOVEMENT how the market got here - cumulative growth and shifting method mix

Market growth

62× growth since 2024

Abliterated and uncensored models in the catalog (delisted ones included) by Hugging Face creation month, cumulative. From 3 in Mar '23 to 21.1K today.

05.7K11.4K17.1K22.8K 3 models by end of Mar '23 21.1K models by end of Nov '26 16K models by end of Aug '26 Mar '23Jan '24Jan '25Jan '26
Each point is the end-of-month total. Includes models later delisted - this is the market as it existed, not just what remains.

Method share by releases

Percentage of the models created on Hugging Face in each complete calendar month, grouped by classified method, counting only labels with medium or high confidence. Reflects what authors are publishing. Currently dominant: M1   Direct removal (36%).

0%25%50%75%100% M1 Direct removal - 36% of most recent month M8 Repackaging (quantization) - 34% of most recent month M3 Layer-wise ablation - 29% of most recent month M1 Direct removal M8 Repackaging (quantization) M3 Layer-wise ablation Apr '24JulJan '25JulJan '26Jul
M1 Direct removal 36%
M8 Repackaging (quantization) 34%
M3 Layer-wise ablation 29%

Supply vs demand gap

Both columns show percentages for the same group of models - those created on Hugging Face in the last complete calendar month, with a medium- or high-confidence method label - but they measure different things: what authors publish (share of those releases) versus what users actually download (share of the lifetime downloads of those same releases). Side by side they reveal which methods are oversupplied, which are undersupplied, and which are outright ignored.

Method % of releases % of downloads Gap Verdict
M1 Direct removal
35.9%
3.5%
-32.4pp Oversupplied vs demand
M3 Layer-wise ablation
28.9%
46.7%
+17.8pp Users want more of this
M8 Repackaging (quantization)
34.5%
49.7%
+15.2pp Users want more of this

Snapshot: models created in 2026-10 (last complete month), medium or high confidence labels only; downloads are the lifetime downloads of the 2026-10 releases to date. Auto-recomputed every hour.

6-month trend

The gap table shows one month. This one shows movement across the last 6 complete calendar months: for each month, each method's share of the lifetime downloads of the models created that month (medium or high confidence labels), and how that share moved from one month's releases to the next. Change is the percentage-point shift from the first month to the last; volatility is grouped into four bands by standard deviation: Low (< 5pp), Medium (5-15pp), High (15-30pp), Extreme (≥ 30pp).

Method Range (min-max) Volatility Change Trend
M8 Repackaging (quantization) 11.2% - 63.7% High -8.8pp ↓ Falling erratically
M3 Layer-wise ablation 21.3% - 86.1% High +8.8pp ↑ Rising but choppy
M1 Direct removal 2.5% - 48.4% High +0.9pp = Sideways with turbulence

Based on the last 6 complete months of releases, weighted by their lifetime downloads. Auto-recomputed every hour.

Full verdict matrix (12 categories = 3 directions x 4 volatility bands)
Low (<5pp) Medium (5-15pp) High (15-30pp) Extreme (≥30pp)
Rising Rising steadily Rising cleanly Rising but choppy Rising through turbulence
Falling Steady decline Clear decline Falling erratically Collapsing
Sideways Locked-in position Sideways, minor noise Sideways with turbulence No trend, wild swings

Method share by downloads

Same months and the same medium- or high-confidence labels, weighted by the lifetime downloads of each month's releases instead of their count. Reveals what users actually run: a few popular methods dominate demand, even if repackaging dominates supply. Currently dominant: M8   Repackaging (quantization) (50%).

0%25%50%75%100% M8 Repackaging (quantization) - 50% of most recent month M3 Layer-wise ablation - 47% of most recent month M1 Direct removal - 3% of most recent month M3 Layer-wise ablation Feb '24JulJan '25JulJan '26Jul
M8 Repackaging (quantization) 50%
M3 Layer-wise ablation 47%
M1 Direct removal 3%
LANDSCAPE what exists in the market - families, sizes, types, authors

Model families

Qwen leads at 44%. Top 3 hold 81% of the family-tagged market.

Forks and lineages

Three views. Generation distribution shows how deep the family trees run: first-generation was built on a base outside the catalog, second-generation is a fork of an abliteration, and deeper means the chain kept going. Most-forked models ranks any parent that gets used repeatedly - foundation models and popular in-catalog abliterations mixed together. Foundation bases only narrows to external upstream weights (Qwen, Llama, Gemma and so on) so you can see which raw material the field prefers, separately from which abliterations become popular re-bases themselves.
EXPLORE the full fork graph →

Generation distribution

orphans
5,725
27.1%
no listed parent
gen 1
7,189
34.1%
built on external base
gen 2
7,450
35.3%
fork of an abliteration
gen 3
656
3.1%
fork of a fork
gen 4
68
0.3%
depth 4
gen 5
5
0%
depth 5

Most-forked models any parent, foundation or in-catalog

01 Qwen/Qwen3.8-27B external 401 forks
02 Qwen/Qwen3.6-35B-A3B external 174 forks
03 Qwen/Qwen-Image-2.1 external 142 forks
04 Qwen/Qwen3.5-9B external 119 forks
05 Qwen/Qwen3.5-35B-A3B external 117 forks
06 google/gemma-4-26B-A4B-it external 111 forks
10 Qwen/Qwen3.6-27B external 100 forks

Foundation bases only external models not in our catalog

01 Qwen/Qwen3.8-27B 401 built on it
02 Qwen/Qwen3.6-35B-A3B 174 built on it
03 Qwen/Qwen-Image-2.1 142 built on it
04 Qwen/Qwen3.5-9B 119 built on it
05 Qwen/Qwen3.5-35B-A3B 117 built on it
06 google/gemma-4-26B-A4B-it 111 built on it
08 Qwen/Qwen3.6-27B 100 built on it
10 google/gemma-4-31B-it 83 built on it
11 Qwen/Qwen3.8-Flash-Next 82 built on it
12 google/gemma-4-12B-it 81 built on it

Evolution chains

The longest ancestor-to-descendant lineages in the catalog. Each arrow is a model that took the previous one as its base and modified it further. About 53 percent of the catalog models that declare a parent are second-order or deeper - this is the deep end.

Sizes

Sweet spot is small 3-10B at 36% of sized models. Giants (80B+) are only 1.7%.

Types

Distribution by capability and format. Multimodal, Mixture-of-Experts, image generation, plus how many are packaged as GGUF for local inference.
GGUF format 10,209 48.4%
multimodal 4,266 20.2%
Mixture-of-Experts 3,012 14.3%
image generation 391 1.9%

Quantization landscape

Distribution of model files by their quantization level. Smaller quants pack the same weights into fewer bits at the cost of some quality; the mix tells you what the community actually deploys.
Q3_K 13,583 files 111244.3 GB
Q4_K 11,964 files 126020.8 GB
Q5_K 10,520 files 121822.9 GB
IQ3 8,384 files 67550.2 GB
IQ2 6,956 files 44072.9 GB
IQ4 6,512 files 61537.8 GB
Q2_K 6,264 files 43444.5 GB
Q6_K 6,057 files 82000.3 GB
Q8_0 5,095 files 72860 GB
Q4 4,178 files 36057.4 GB
F16 3,532 files 28253.5 GB
IQ1 2,953 files 13013.4 GB
BF16 2,672 files 38452 GB
Q5 587 files 4413.3 GB
Q8_K 454 files 14634.9 GB
F32 255 files 1150.2 GB
FP8 124 files 1056.7 GB
GPTQ 3 files 15.6 GB

Licenses

apache-2.0 dominates at 62% of models that declare a licence.

Authors

Two views of the 5,547 people publishing to this catalog: Volume ranks who ships the most, Engagement ranks who actually maintains their releases in discussions. The header numbers describe the whole author base regardless of which view is open.
0.674
Gini coefficient
moderate concentration
17.1%
Top author's share
22.4%
Top 5 authors' share
26.6%
Top 10 authors' share

Four indices of the frontier

READ methodology →
Four metrics we are establishing to track how this niche evolves over time. Two are measurable today, and one of them is rising. Two hold at zero because no confirming incidents have been publicly documented yet. Publishing them at zero is the point - we want a permanent record of the day the numbers moved.
Last 24 h 53 new models, by Hugging Face creation date · Last 7 days 490 · Year to date 14,302
112.5 days
Freedom Velocity Index
Average days from a base model's release on Hugging Face to the first abliterated or uncensored model built on it, across 480 original base models released since June 2024. Shorter = faster niche.
0.0%
Weaponization Index
Share of abliterated models publicly linked to real-world harm. Zero verified cases as of the latest review, 2026-11-11.
N/A
Freedom-to-Weapon Latency
Days from the first abliterated or uncensored model built on a base model to its first documented misuse. Not measurable while the Weaponization Index reads zero, because there are no incidents to measure from.
19.1 %
Disenchantment Index
Share of a base model's downloads that flow into abliterated descendants at any tree depth. Weighted across 534 base models.
↗ 26mo trend
For every base model with at least one abliterated descendant (direct forks + forks-of-forks recursively) we compute what share of its download market has been "disenchanted". The market-weighted value below is the one to cite: total descendant downloads divided by total base downloads across all 534 of those bases. The top-10 table further down is narrower: it ranks only bases with at least 3 descendants and at least 1,000 base downloads, a floor that filters out obscure bases whose tiny denominator would push DAI into the tens of thousands of percent - divide-by-noise, not a market signal.
19.14%
of the total open-model download market has flowed into abliterated descendants
Up from 0.15% in 2024-10 - 124.4x growth over 24 months. Recent 6-month growth: 7.4x.
MARKET DAI, 26 MONTHLY SNAPSHOTS
0.0% 5.3% 10.5% 15.8% 21.1% 2024-10: 0.15% 2024-11: 0.18% 2024-12: 0.23% 2025-01: 0.32% 2025-02: 1.66% 2025-03: 0.53% 2025-04: 0.39% 2025-05: 0.38% 2025-06: 0.38% 2025-07: 0.50% 2025-08: 0.64% 2025-09: 0.65% 2025-10: 0.40% 2025-11: 0.42% 2025-12: 0.61% 2026-01: 0.61% 2026-02: 0.63% 2026-03: 1.27% 2026-04: 2.20% 2026-05: 2.60% 2026-06: 3.45% 2026-07: 2.89% 2026-08: 6.41% 2026-09: 7.21% 2026-10: 11.62% 2026-11: 19.14% 2024-102025-012025-042025-072025-102026-012026-042026-072026-102026-11
Weighted DAI at each of the 26 monthly snapshots. Peak reading so far: 19.14% in 2026-11. Line rising = the abliterated market is growing faster than the base market.
GROWTH RATE - HOW FAST DAI IS CHANGING
-77% 143% 363% 583% 803% 2025-01: 105.1% growth over prior 3 months 2025-02: 803.4% growth over prior 3 months 2025-03: 134.4% growth over prior 3 months 2025-04: 23.3% growth over prior 3 months 2025-05: -77.2% growth over prior 3 months 2025-06: -28.6% growth over prior 3 months 2025-07: 28.3% growth over prior 3 months 2025-08: 68.5% growth over prior 3 months 2025-09: 71.7% growth over prior 3 months 2025-10: -19.7% growth over prior 3 months 2025-11: -33.4% growth over prior 3 months 2025-12: -6.6% growth over prior 3 months 2026-01: 52.9% growth over prior 3 months 2026-02: 47.3% growth over prior 3 months 2026-03: 107.4% growth over prior 3 months 2026-04: 258.4% growth over prior 3 months 2026-05: 316.0% growth over prior 3 months 2026-06: 172.5% growth over prior 3 months 2026-07: 31.4% growth over prior 3 months 2026-08: 146.2% growth over prior 3 months 2026-09: 109.0% growth over prior 3 months 2026-10: 302.6% growth over prior 3 months 2026-11: 198.5% growth over prior 3 months 2025-012025-042025-072025-102026-012026-042026-072026-102026-11
3-month rolling growth rate of DAI: for each point, how much DAI has changed relative to 3 months earlier. Rising line = the market is accelerating. Falling line = growth is slowing. Values below zero mean DAI is shrinking.
TOP 10 MOST DISENCHANTED BASES (3+ DESCENDANTS, 1,000+ BASE DOWNLOADS)
Base model Base downloads Descendants Descendant downloads DAI
Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B 3,440 86 1,708,279 49659.3%
Accio-Lab/occamy-1.0 1,428 5 63,903 4475.0%
LiquidAI/LFM2.5-24B-A2B 27,980 16 1,062,457 3797.2%
empero-ai/Qwen3.8-35B-A3B-Distill 3,821 4 127,391 3334.0%
Aleph-Alpha/Kolibri-1-BF16 2,646 5 63,320 2393.1%
empero-ai/Qwythos-9B-Claude-Mythos-5-1M 3,659 26 68,599 1874.8%
bottlecapai/ThinkingCap-Qwen3.8-27B 2,583 9 32,379 1253.5%
unsloth/Phi-3-mini-4k-instruct 1,606 8 18,948 1179.8%
ukisai/Swift-1.5-Qwen3.8-27b 5,260 23 59,729 1135.5%
Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2 1,581 16 16,574 1048.3%

The 1% rule

A tiny fraction of models absorbs almost all the demand. Every abliterated model that is not delisted and has at least one lifetime download is ranked (lifetime = the Hugging Face all-time count from the hub-stats snapshot, or the last-30-days counter when that is higher), then we ask what share of total lifetime downloads the top slice controls. Result: an extreme power law, sharper than most consumer markets.
63.4%
of all lifetime downloads go to the top 1% of models
that is 182 models out of 18,220
81.9%
top 5% (911 models)
88.5%
top 10% (1,822 models)
94.1%
top 20% (3,644 models)
0.9%
bottom half (9,110 models) share this

How much of the catalog is still alive

Cataloged doesn't mean used. Every model here has a live counter from HuggingFace showing downloads over the last 30 days - the same window HuggingFace itself uses. Here is the honest activity distribution: how many models are still getting real traffic, how many are barely alive, how many are outright abandoned.
2,869
actually active
14.5% get 1K+ downloads / month
9,783
have real traffic
49.4% get 100+ downloads / month
10,034
effectively forgotten
50.6% get less than 100 / month
FULL DISTRIBUTION (LAST 30 DAYS)
10K+ per month
519 2.6%
1K - 10K
2,350 11.9%
100 - 1K
6,914 34.9%
10 - 100
6,497 32.8%
1 - 10
1,814 9.2%
0 downloads
1,723 8.7%

19,817 models that are not delisted (the catalog holds 21,093 including delisted ones). Numbers refresh hourly from HuggingFace live counters.

Original vs derivative work

19,817 models that are not delisted sounds like a big catalog. It is - but not all of it is what most people would call "original work". Direct abliteration and fine-tune methods (M1, M3, M4, M-uncensored) genuinely change model weights or behavior. Repackaging (M8) converts existing models into different quantization formats. Merges (M5) combine already-abliterated models via arithmetic. Here is the honest split.
11,382
57.4%
Original abliteration or fine-tune
Actual weight surgery or preference training (M1, M3, M4, M-uncensored). This is where new abliterated models come from.
5,939
30%
Repackaged (GGUF conversions)
M8. Existing abliterated models re-quantized for llama.cpp / Ollama / LM Studio. Necessary work, but no new abliteration.
682
3.4%
Merged (mergekit)
M5. Weight arithmetic combining several already-abliterated models. Second-order production.
1,814
9.2%
Not yet classified
Metadata insufficient for the classifier. Coverage improves as we index more.
FULL BREAKDOWN BY METHOD
Method Models Share of catalog Share of lifetime downloads
M8 Repackaged (GGUF) 5,939 30% 33%
M-uncensored Uncensored fine-tune 4,353 22% 29.5%
M1 Direct removal (abliteration) 4,348 21.9% 8.4%
M3 Layer-wise ablation 2,577 13% 21%
unknown Not yet classified 1,814 9.2% 4.5%
M5 Merged (mergekit) 627 3.2% 0.7%
M4 Hybrid (abliterate + heal) 104 0.5% 2.1%
M5-frankenstein Frankenmerge 55 0.3% 0.8%

Demand-side inequality (Gini)

Our other Gini measures who produces models. This one measures who receives the downloads: lifetime downloads of every model that is not delisted and has at least one. Both matter, and they usually differ: producers concentrate around a few prolific pipelines, but demand can concentrate even more sharply on a handful of proven models. A value of 1 means one model gets all downloads; 0 means downloads are spread equally.
0.921
Gini coefficient by downloads
EXTREME CONCENTRATION
Models measured 18,220
Total lifetime downloads pool 280,040,348
For reference: US household income Gini is roughly 0.48. Values above 0.9 are rare outside heavy-tailed digital markets like open-source packages or viral content.

Context length distribution

Context sizes cluster at powers of two (128K, 256K) and at a nominal 1M (2^20 tokens). Almost no models sit in the 256K-to-1M range - it's a real gap. And a handful of Llama-4-Scout variants reach all the way to 10M. Based on 7,955 models with metadata (GGUF or safetensors).
up to 4K 297 3.7%
4K - 32K 1,658 20.8%
32K - 128K 2,821 35.5%
128K - 1M 2,870 36.1%
1M 302 3.8%
10M (Scout) 7 0.1%

Method firsts

For each classified method, the earliest model on record. Not "who invented the technique" - that's often academic - but who published the first release we can classify with confidence.

Disappeared models

EXPLORE all 3,338 deletions →
3,338 models in this niche have vanished from HuggingFace since October 2024 - 13.7% of the current catalog plus the deletions. Models still live in our catalog are not counted. Average lifespan before deletion: 7.4 months. 2 had over a million lifetime downloads before disappearing. Nowhere else public tracks this.

Disappearances by month

24
10/24
23
11/24
20
12/24
30
01/25
36
02/25
20
03/25
26
04/25
92
06/25
13
07/25
22
08/25
324
09/25
214
10/25
30
11/25
112
12/25
51
01/26
1056
02/26
133
03/26
77
04/26
149
05/26
176
06/26
201
07/26
205
08/26
304
09/26
Most significant disappearances (by lifetime downloads)
abenzerps/Qwen-Image-2.1-Uncensored-GGUF 1,152,523 lifetime · 1mo alive last seen 2026-09
LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V12-GGUF 1,005,078 lifetime · 1mo alive last seen 2026-08
LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Wasserstein-GGUF 462,887 lifetime · 3mo alive last seen 2026-06
LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-GGUF 444,251 lifetime · 1mo alive last seen 2026-09
huihui-ai/Llama-3.3-70B-Instruct-abliterated-finetuned-GPTQ-Int8 437,999 lifetime · 10mo alive last seen 2025-09

Refusal direction extraction methods

How the refusal direction was found for each model where we can determine it. The M1-M8 classifier tags what was done to the weights; this shows the geometric method behind that operation. Coverage today: 7,453 of 21,093 models (35.3%). 7 methods documented in the literature; 2 are visible in production.
Used in production
Difference-of-means
7,264 (34.4%)
huihui-ai layer-band extraction
189 (0.9%)
5 documented in literature, no production adoption: Affine Concept Editing (ACE)·Contrastive Activation Addition·PCA/SVD on activation differences·Recursive Feature Machines with AGOP·Self-Organizing Maps
Theory has run ahead of practice. 5 documented extraction methods exist in academic literature but have not yet been observed in released models. The distribution above is one useful angle on the state of the field: rich theoretical vocabulary, narrow production practice. Read the full analysis in the wiki →

Curious how these numbers are calculated? Every metric on this page is a plain SQL query against the catalog database. The full methodology - every definition, every formula, every caveat - is documented in the wiki. Read: every number on the market page, explained →

Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration