Hosting landscape: who distributes and serves safety-removed weights
Not a recommendation list.
Weight publishers (Hugging Face)
| Publisher | Pattern | Relevance |
|---|---|---|
| huihui-ai | Ablated builds of nearly every major open release (Qwen, Gemma, Llama, GLM, gpt-oss, VL models) within days; also on Ollama as huihui_ai/* |
The default source in the wild; two of our routes come from here |
| mlabonne | The 2024 reference ablations and the DPO-healing recipe | Historical reference point; less active on 2026 releases |
Community -heretic builds (e.g. acc100) |
Heretic-optimised, low KL | Our Spark 27B comes from here |
| cognitivecomputations | Dolphin fine-tunes | Fine-tune lineage, not ablation |
| NousResearch | Hermes 4, neutral alignment | A serious lab, not a hobby publisher; policy by system prompt |
| Quantisers (unsloth, cyankiwi, QuantTrio, RedHatAI, lhca521, bartowski) | AWQ / NVFP4 / GGUF repacks of the above | Where runnable artifacts come from; the quant changes refusal behaviour on its own |
Time-to-ablation for a new open release is now days. Track it per release in the news log; it is the single best indicator of how pointless weight-level defenses are.
API providers
| Provider | What they host | Retention claim | Notes |
|---|---|---|---|
| Venice.ai | Open models plus own -uncensored fine-tunes on a private tier; proxies closed models on an anonymised tier where content is still visible upstream |
Contractual ZDR on private tier | Our venice-deepseek-v41-uncensored route; house system prompt applies; OpenRouter partner |
| Featherless.ai | ~53 000 HF models served directly (2026-10-27 count) with explicit ablated and uncensored categories | Says no prompt storage | Deepest catalog |
| OpenRouter | Aggregator; carries the Venice Dolphin edition; free-tier status flipped during 2026 | No logging unless opted in | Convenient for A/B |
| DeepInfra, Together, Hyperbolic | Some community fine-tunes on demand | Varies | |
| Abliteration.ai | Ablated-only API, new in 2026 | Says no storage [unverified] | |
| xAI | Not ablated; permissive policy | xAI AUP | The "permissive frontier" comparison point |
What stays enforced everywhere
An acceptable-use policy at the account layer, the model's own residual refusals, and rate limits. "Uncensored model" is not "uncensored account". All retention claims are provider statements, not verified.
Sources
venice.ai/blog (uncensored APIs, OpenRouter partnership) · subconscious.dev Venice review 2026 · experientiallabs.ai (2026-10) · deai.org refusal policies compared · freellm.net, free-model.com OpenRouter listings · featherless.ai blog · HF publisher pages. Vendor-authored; [unverified].