Disenchantment Index (DAI)

What DAI measures, in one paragraph

For any base model - a foundation LLM published by an organization like Alibaba, Google, Meta, Mistral - the Disenchantment Index measures what portion of its total download market has gone into abliterated descendants rather than into the original. Descendants include every fork, every fork of a fork, every quantization of a merge of a fine-tune of an abliteration, traced through the fork tree to any depth. If Qwen3.8-27B has been downloaded 5.25 million times and the sum of downloads of every abliterated model in its recursive fork tree comes to 10.63 million, its DAI reads 202% - the abliterated audience has become larger than the audience for the base itself.

Aggregated across every foundation model in the ecosystem with at least three abliterated descendants, the market-wide DAI currently reads 8.26%. That number - roughly one download in 12 across the tracked base of 478 foundation models - is the flagship reading of this index. In October 2024 the same reading was 0.15%. The ratio has grown by 53.6x in twenty-four months.

The formula, spelled out

For a single base model, let B be the total downloads of the base at time t, and let D(B, t) be the sum of downloads at time t across every model in the recursive fork tree rooted at B. Then:

The recursive traversal matters. A direct fork is one whose yaml_base_model field names B. But a great deal of ecosystem activity happens two, three, or four levels below the base: someone abliterates Qwen; someone else quantizes that abliteration into GGUF for llama.cpp; a third party merges that GGUF with another abliteration; a fourth quantizes the merge. Every download at every depth expresses the same underlying choice - the choice to consume an audience that has been stripped of its refusal layer. Counting only the direct children would understate the phenomenon by a factor of three to five for most popular families.

Aggregated to the market level, the weighted DAI is:

This is the same aggregation the S&P 500 uses for its constituents - market-capitalization-weighted rather than a simple average. It gives foundation models weight in proportion to their audience size. A giant model like BERT-base contributes to both numerator and denominator, but its low DAI value pulls the market number toward reality: most people who download BERT are not there for its abliteration. A middle-weight model like Qwen3.8-27B with a healthy abliterated ecosystem contributes a large positive term. A tiny model with an outlier DAI in the thousands of percent contributes almost nothing because its base downloads are almost nothing.

This is why we do not report a simple arithmetic mean of per-base DAIs. That statistic - "the typical base has a DAI of X" - is dominated by long-tail outliers (small models where a few enthusiast forks produce a wildly larger audience than the original). The market-weighted number is the honest one for the ecosystem as a whole; the per-base numbers are honest at their own scale.

Where the name comes from

A language model as its lab ships it comes with a moral coating - a layer added during training that makes it refuse certain prompts and hedge others. Abliteration is the operation that removes this coating: what was framed as a boundary turns out to be a direction in activation space that can be identified and subtracted. The name of this index gestures at the German sociologist Max Weber's phrase Entzauberung der Welt - the disenchantment of the world - because the operation looks analogous: a thing framed as sacred and non-negotiable becomes an editable object. The reference is etymological, not philosophical. We are counting download flows, not building a theory.

The Disenchantment Index measures how much of the world has actually chosen this transformation. If DAI were zero everywhere, we would say that abliteration exists but no one wants it. If DAI were 100% everywhere, the safety layer would have become optional - shipped by the vendor, not preferred by the audience. The real number is between these extremes and rising. That rise is what this index tracks.

Reading the market number

The current market reading is 8.26%. Put differently: for every 12 downloads across the tracked foundation models in aggregate, one download goes to an abliterated version. That is a number worth sitting with. Two years ago it was one in 667. Six months ago it was one in 38.

The number will not stay put. Every day the automation refreshes the underlying snapshots and recomputes DAI from scratch. New abliterations enter the tree; obsolete ones get deleted from HuggingFace and drop out. If a large new base model is released with poor safety-tuning, we would expect DAI for that family to spike quickly; if a base model is quietly deprecated and replaced with a version that satisfies users without needing abliteration, its DAI would decline. This index will move as the market moves.

Reading a per-base number

A base model's DAI can and often does exceed 100%. This is not a mistake in the calculation and it is not a defect in the data. It is the signal we most care about. A DAI of 200% for a base means the sum of downloads of its abliterated descendants is twice the downloads of the base itself - the audience for the disenchanted version has grown larger than the audience for the original.

Among the biggest names in the current top of the table: Qwen/Qwen3.8-27B reads 196% (6,758,884 downloads on the base, 13,272,511 on its 708 descendants). Each of these is a foundation model published by a large organization, released with commercial-grade safety tuning, and audibly under-consumed relative to its abliterated tree. The ratios are visible facts on HuggingFace's download counters.

The extreme tops of the table - values in the thousands of percent - belong to a different kind of story. There the base is a small research checkpoint, perhaps a mid-training snapshot published for reproducibility, with a few hundred downloads to its name. Its abliterated descendants sit in the tens of thousands of downloads. The ratio is real; the interpretation is different. These are cases where a niche foundation model was rescued from obscurity by its abliteration - the fork became the point of the family. Not the majority of the table, but the most philosophically striking rows in it: the disenchanted version is not the choice of the audience relative to the original; it is what the audience discovered exists.

Two waves

The 24-month history of the market-wide DAI reads clearly, once you look at it. From October 2024 through January 2025, the reading sat between 0.4% and 1.0% - a niche existed, but marginal. In February 2025 a spike to nearly 4% arrived - the release of Qwen 2.5 and the maturation of Heretic, the first widely-used tool for pushbutton abliteration, both hit that month, and both left a mark. The spike did not hold. For roughly the following year the market receded to under 1%.

The second wave has been unfolding through 2026. Six months ago DAI read 2.61%; today it reads 8.26%. That climb, in pure ratio terms, is a 3.2x acceleration over six months - larger than the entire trajectory of the year preceding it. What triggered it is a question this index does not answer directly, but the factual claim it makes is unambiguous: we are, right now, in the middle of the largest expansion of the abliterated market share ever observed in the two-year window this data covers.

A rising line does not, of course, guarantee that it keeps rising. The February 2025 spike also looked at the time like a beginning, and it turned out to be an isolated event followed by a return to baseline. Only future readings will show whether the current wave is a permanent shift in the ecosystem or another spike. That is precisely why the index exists: to make the readings visible, so that this question becomes answerable.

The bases where the movement is strongest

Below is a snapshot of the top of the ranking as of 2026-10. It is generated live from the same query that populates the detail page and refreshes daily with the underlying data.

1. lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled
   Base downloads: 5,665. Descendants: 46 models totaling 173,811 downloads. DAI: 3068.2%.
2. Qwen/Qwen-Image-2.1
   Base downloads: 94,556. Descendants: 59 models totaling 1,809,624 downloads. DAI: 1913.8%.
3. ukisai/Swift-1.5-Qwen3.8-27b
   Base downloads: 4,292. Descendants: 9 models totaling 53,135 downloads. DAI: 1238.0%.

The complete ranking, sortable and searchable, with per-base 24-month sparklines and expandable full history charts, lives at /market/disenchantment. Each row on that page also carries an ACTIVE or INACTIVE badge based on whether the base's descendant downloads have grown or contracted over the last three months, and a link to browse all of that base's forks in our catalog with one click.

What DAI does not measure

Quality. A high DAI does not mean the abliterated versions are better than the base. It means they are downloaded more. Quality signals - benchmark scores, community feedback, actual usage patterns - require different instruments and are not captured here.

Intent. The index counts downloads, not the reasons for them. A researcher studying refusal behavior downloads an abliterated model for reasons that are indistinguishable, at the counter level, from a hobbyist who wants an unhinged roleplay partner or a security researcher red-teaming a system. DAI is agnostic to intent.

Harm. DAI is not a safety metric. A rising DAI does not mean more real-world harm is happening; a falling DAI would not mean it is decreasing. Those are questions for the sibling Weaponization Index, which counts documented incidents rather than market flows.

Adoption in production. HuggingFace downloads are download events - a person or a script pulled the weights. What happens after is not visible from download counts. A model downloaded ten thousand times may sit unused on ten thousand hard drives, or it may serve millions of API calls. DAI is a demand-side signal, not a deployment signal.

Open methodological questions

Base filter boundaries. We currently exclude from the base list any model whose ID contains recognizable abliteration keywords (abliterated, uncensored, heretic, jailbreak, dolphin, unfiltered, decensored) or that already exists in our own catalog. This prevents double-counting - we do not want an already-abliterated model to appear as a base whose "descendants" produce inflated DAI. But some models slip through - a fine-tune with a neutral name that in fact removes some safety behavior, indistinguishable at metadata level from a legitimate foundation model. These cases add noise to the top of the ranking. Extending the filter to include behavioral signals (empirical refusal-rate testing) is on the roadmap.

The yaml_base_model declaration is honesty-based. Authors self-declare what base their model derives from in the model card frontmatter. Most do so correctly; some leave the field blank; some name an intermediate model rather than the true root; some name a model that no longer exists. The recursive tree walk relies on this declared graph. Mislabeled edges lead to miscounted trees. There is no fully automated way to verify the claim without loading and inspecting the weights themselves; the current index accepts the declared graph and notes this as a known source of noise.

Small-base sensitivity. A base with, say, one hundred downloads and a descendant tree of ten thousand downloads produces a DAI of ten thousand percent. This is mathematically correct and philosophically interesting (the base was overshadowed by the derived work) but it is also fragile - one popular derivative can move the number wildly. We report these rows without cap because we believe the signal is genuine, but we acknowledge that a single outlier fork can dominate a base's reading.

Deletion asymmetry. When an abliterated model is deleted from HuggingFace, it stops contributing to descendant counts on the following snapshot. When a base model is deleted, the whole subtree loses its denominator and the base drops off the ranking entirely. This means the index responds asymmetrically to deletion events - a topic we track separately in the deletion history and plan to correlate with DAI trajectory in future analysis.

Where the numbers come from

Snapshots of the entire HuggingFace model registry are collected monthly from the cfahlgren1/hub-stats dataset - a public, maintained dump of the model index with download counts. Twenty-four consecutive monthly snapshots covering October 2024 through the current month sit on our server (roughly 19 GB of Parquet). Our own catalog of abliterated and uncensored models sits alongside, indexed continuously.

Every night at 05:00 UTC an automation pulls the latest snapshot, recomputes every base's recursive descendant tree, updates the base-model history and market history tables, and rebuilds the site. The numbers you see on this page and on the detail page are never more than 24 hours old.

The full computation pipeline is open. The Python scripts that build the tables (enrich_base_models.py, build_dai_history.py) live on our server and can be inspected. The queries that read them into the site (getDisenchantmentIndex, getDaiMarketHistory) are part of the site codebase.