Every number on the market page, explained

Why this article exists

Every number on the market page is the answer to a plain question. "How many models exist?" "Who publishes the most of them?" "How fast does the community react when a new base model drops?" "What share of the market has flowed into abliterated versions?" Behind each answer sits a SQL query and a definition. Both matter, and both are documented here.

Nothing is proprietary. The catalog database is not yet publicly downloadable (that ships with the API we are building), but every definition and every SQL formula lives in this article or in a linked one. If a number looks wrong, disputes are welcome by email.

The catalog is the foundation. 18,391 indexed models is what makes anything else possible - the analytics on this page, the search and filtering, the app suggestions, this wiki. Every metric below is a specific question asked of that catalog.

How the market page is organized

Three tabs sit at the top. Each has a different purpose:

• Overview - the default view. Headline totals, family shares, community reaction time, author concentration, current momentum, model of the month, evolution chains, quantization landscape, timeline, and refusal-direction extraction methods. This is the "state of the market right now" tab.
• Indices - the flagship metrics that track how the niche evolves. Four of them: Freedom Velocity, Weaponization, Freedom-to-Weapon Latency, and the Disenchantment Index. Two are measurable today; two hold at zero by design and will move when the first documented incident lands. See the dedicated wiki articles: three frontier indices and Disenchantment Index.
• Signals - structural signals about how the market behaves. Pareto concentration, activity breakdown, originality vs derivative work, demand-side Gini, context length distribution, method firsts, disappeared models, extraction methods.

Two extra pages, linked from within the tabs, hold the biggest datasets in their own place: /market/deletion-history for the full timeline of 2,860 disappearances, and /market/disenchantment for the ranked table of every base model with abliterated descendants.

Overview: the three headline numbers

The top of Overview shows three big numbers. They set the scale for everything that follows.

• Indexed models - total abliterated and uncensored language models in our catalog: currently 18,391. This is the ceiling for every other percentage on the page.
• Distinct authors - unique publisher usernames: currently 4,551. One person can publish under multiple usernames, so this counts usernames rather than people.
• Unique base models - how many different original models the community draws from: currently 3,254. A low number relative to indexed models would mean everyone is working on the same handful; the actual ratio shows source diversity.

Overview: the split hero and the DAI panel

Above the three headline numbers sits a two-column hero. On the left, the growth summary: how many models today, this week, this month, plus the multiplier since 2024. On the right, the flagship Disenchantment Index panel - one number, one line of context, a link to the Indices tab.

Current DAI reading: 8.74%. Up from 0.15% in 2024-10 - a 56.8x shift over the two-year window. Full explanation in the DAI wiki article.

Overview: model families and market shares

Every abliteration is built on top of some base model. That base belongs to a family - Qwen, Llama, Gemma, Mistral. The market page ranks the top families by count.

Current leader: qwen with 6,309 models (0.0% of family-tagged catalog). The ranking shifts month by month as new base models get released and the community reacts.

The percentage denominator is only family-tagged models, so shares add up to 100. If we used the total catalog as the denominator, orphan models with unrecognizable families would eat into every percentage.

Overview: community reaction time

This is the metric we exist to compute. It answers a specific question: when a new family of language models appears in the wild, how fast does the abliteration community react?

For each family, we find the earliest model in the family in the catalog, then the earliest abliteration built from anything in that family, and measure the gap. Faster reaction times mean the family is a popular target. Slower reaction times mean it took a while to catch attention.

The leading families sort by size, not by reaction time - so the row on top has the most abliterations overall, not the quickest response. Reading across each row: first abliteration in X days, Y appeared during the first 30 days, current rate is Z per month, W total in catalog.

No one else publishes this metric. Every leaderboard measures what a model can do. This measures how fast the market reshapes itself around a new arrival. It matters for AI-safety researchers, regulators, and model producers themselves.

Overview: author concentration (supply-side Gini)

How centralized is publishing? Is the market dominated by a few prolific authors, or spread across many? Two numbers answer this: the Gini coefficient and the top-N shares.

Current Gini reading: 0.691. On the 0-1 inequality scale, values above 0.7 signal high concentration; above 0.85, extreme. A companion Gini (on the demand side, computed from downloads rather than production counts) lives on the Signals tab; see the section below.

Top-N shares are cumulative: Top 5 includes Top 1, Top 10 includes Top 5.

Overview: current momentum

Which families is the community actively working on right now, and is the pace picking up or cooling down? Three overlapping windows (30, 60, 90 days) plus a trend arrow that compares the last 30 days against the 30 before.

Arrows: ↑ if the last 30 days produced 10% more than the previous 30, ↓ if 10% less, → if inside a ten-percent band. The band is a judgment call - small enough to catch real acceleration, large enough to filter month-to-month noise.

Overview: model of the month

The single most-downloaded abliteration whose creation date is within the last 30 days. Download counts are cumulative from release, so "most downloaded in the last 30 days" is really "the recently-released model that has attracted the most downloads so far."

Currently: abenzerps/Qwen-Image-2.1-Uncensored-GGUF with 2,188,658 downloads.

Overview: evolution chains and quantization landscape

Half of every abliteration in the catalog is second-order or deeper: someone took an already-abliterated model and modified it further, and often someone else took that and modified it further again. Evolution chains trace those lineages from the original base to the deepest descendant.

Chains shorter than three generations are dropped. The market shows the top eight deepest chains, sorted by depth then by how many descendants the root model has.

Below the chains, the quantization landscape counts files (not models) by quantization level - from tiny 2-bit files that fit on a phone to full-precision files that need a datacenter. Useful for inference providers deciding what to host.

Overview: timeline and extraction methods

A month-by-month histogram of model creation, going back 24 months. Each bar is how many models were created in that month, according to each model's original release date.

Then a section on refusal direction extraction methods - how the refusal direction was found for each model where we can determine it. Seven methods are documented in the academic literature; two are visible in production (Difference-of-means and huihui-ai's layer-band). The other five (Affine Concept Editing, Contrastive Activation Addition, PCA/SVD, Recursive Feature Machines with AGOP, Self-Organizing Maps) exist on paper but have not been observed in a released model yet. Theory has run ahead of practice.

Indices tab

Four flagship metrics that track how the niche evolves over time. Two are measurable today; two hold at zero by design.

• Freedom Velocity Index - the average days from a base model release to its first abliteration. Currently 65.8 days across 19 tracked families.
• Weaponization Index - the share of abliterated models publicly named in documented real-world harm. Currently 0.0%. Zero verified incidents as of the latest review. Publishing at zero is the point: this creates the baseline against which future movement is measurable.
• Freedom-to-Weapon Latency - days between the first abliteration of a family and the first documented misuse of that family. Not measurable while Weaponization is zero.
• Disenchantment Index (DAI) - share of a base model's download market that has flowed into abliterated descendants at any tree depth. Currently 8.74% weighted across 478 bases. Full mathematical treatment in the DAI wiki article.

Below the four-index panel, the Indices tab includes a Disenchantment detail block - the flagship DAI number bigger, plus two 24-month charts (the level of DAI over time, and its rolling growth rate), plus the top-10 most-disenchanted bases. Deeper details at /market/disenchantment.

Signals tab

The Signals tab holds structural signals about how the market behaves. Eight blocks.

Pareto concentration (the 1% rule)

What share of downloads goes to the top 1% of models, the top 5%, and so on. In most healthy markets the top 1% captures around 20-30% of demand. In this one, the top 1% currently captures around 80% - an extreme long-tail distribution where a few models absorb almost all attention. The Signals block gives the number and interpretation.

Activity breakdown - how much of the catalog is still alive

Every model has a live counter for downloads over the last 30 days. The activity block classifies the catalog into six buckets. Right now: 2,853 get 1K+ downloads/month (active), 9,145 get 100+ (traffic), and 8,486 get less than 100 (effectively forgotten). The full six-bucket distribution appears below the hero cells.

Original vs derivative work

A breakdown of the catalog by whether each model represents new weight surgery (M1, M3, M4, M-uncensored), a re-quantization of existing work (M8 GGUF), a merge (M5), or unclassified. Roughly 54% is original abliteration and 32% is GGUF repackaging. Necessary distinction - not every model in the catalog is fresh production.

Demand-side inequality (Gini)

A second Gini coefficient, computed on downloads rather than on production. Producers concentrate around a few prolific pipelines; demand concentrates even more sharply on a handful of proven models. Current reading: 0.939 - extreme concentration. For reference, US household income Gini is roughly 0.48; values above 0.9 are rare outside heavy-tailed digital markets.

Context length distribution

How model context sizes cluster. Context lengths follow a power-of-two pattern (128K, 256K) with a nominal 1M ceiling and a handful of Llama-4-Scout variants reaching 10M. Almost no models sit in the 256K-to-1M range - a real gap in the market. Based on models with GGUF or safetensors metadata.

Method firsts

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

Disappeared models

2,860 models in this niche have vanished from HuggingFace over the last 24 months - roughly 15% of everything ever indexed. Average lifespan before deletion: about 8 months. Some had over a million lifetime downloads. The Signals block shows a monthly timeline of disappearances plus the most significant deletions by lifetime downloads. Full 2,860-row table at /market/deletion-history.

Extraction methods (production and theory)

A second treatment of extraction methods, this time in the Signals context: how many models use each documented technique in production versus how many exist only in academic literature. Two methods are visible in production; five are documented but not observed. Rich theoretical vocabulary, narrow production practice.

Extra pages

Two datasets are big enough to deserve their own page.

• /market/deletion-history - the full 2,860-row disappearance table. Sortable by lifetime downloads, months alive, or deletion date. Searchable. Includes the monthly timeline chart in bigger form.
• /market/disenchantment - the full ranked table of every foundation model with three or more abliterated descendants (currently over 450). Each row includes a 24-month sparkline; clicking a row reveals the full history chart with base downloads and descendant downloads as two lines. ACTIVE / INACTIVE badge based on recent trend; "N forks" link opens the catalog filtered to that base's tree.

Where to read more

Dedicated wiki articles that go deeper on specific parts of the market page:

• Three frontier indices - Freedom Velocity, Weaponization, Freedom-to-Weapon Latency in detail
• Disenchantment Index (DAI) - the full mathematical treatment plus etymology and open methodological questions
• The fork graph - how the family tree of abliterated models is built
• Academic research - 55 arXiv papers organized by attack, theory, methodology, defense, evaluation

Checking anything you see

Once the API ships, running these queries against a live copy of the database becomes possible from anyone's terminal. Until then, disputes are welcome by email. Aggressive scrutiny of individual numbers is not a threat - it is exactly what a serious market analytics resource is supposed to attract.