By UNST33R
The abliterated model ecosystem, by the numbers (October 2026)
A snapshot of the abliterated ecosystem on Hugging Face: 100 checkpoints, 8.7M downloads, and why Qwen dominates the base-model split. Data pulled October 22, 2026.
If you run red-team harnesses, the abliterated ecosystem is your supply chain. Here is what it actually looks like on Hugging Face right now, from a query pulled October 22, 2026 (huggingface.co/api/models?search=abliterated, sorted by downloads).
The headline numbers
- 100 checkpoints in the top search results alone
- 8.69M combined downloads across those checkpoints
- 11,064 combined likes
- The single most-downloaded checkpoint (huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF) has 2.39M downloads and 1,453 likes on its own
For comparison, the broader "uncensored" tag (fine-tunes, merges, and ablations mixed) shows the same shape: the top entry is also a Qwen3.8-27B derivative at 2.4M downloads.
The base-model split
| Base family | Share of top-100 checkpoints |
|---|---|
| Qwen | 49% |
| Other (GLM, Phi, custom merges) | 33% |
| DeepSeek | 8% |
| Gemma | 6% |
| Llama | 4% |
Two things worth reading out of this. First, Qwen is the default substrate for abliteration work — dense variants, permissive licensing, and strong baseline capability make it the path of least resistance. Second, Llama's share has collapsed to near-irrelevance in this niche; the abliteration community migrated.
What operators should note
- GGUF dominates the distribution format. Nearly every high-download checkpoint ships GGUF first. If your harness only speaks safetensors endpoints, you are out of sync with where the ecosystem actually is.
- "Healing" variants are appearing — abliterated checkpoints with post-ablation fine-tuning to recover tool-call JSON fidelity. If strict function-calling matters in your loop, prefer these or validate before relying on the raw projection.
- Download counts ≠ recency. Several 2M+ checkpoints were published months ago. For refusal-behavior baselines, check the base model version behind the ablation, not just the download rank.
Running these against a real endpoint
Self-hosting a GGUF drop is the right move for local experiments. When you need the same families behind an API for harness runs, the UNST33R catalog serves unst33r-deepseek (abliterated DeepSeek-V4.1-Flash) and uncensored GLM-5.3 tiers per token — the tier comparison maps them to workloads.
Method note: counts are from a single API pull and will drift; the underlying pages are the source of truth. As always — authorized testing only, per the acceptable-use policy.