Qwen3.8-27B · Ablated
Available nowOur first release. The refusal direction is removed directly from the base weights — no retraining, no degraded reasoning. It answers what you ask, not what a filter allows.
- 27B dense
- Zero refusals
- Open weights
More precise · Less damage
Remove the model's censorship
Keep its original capabilities
Ablatehub hosts models with their censorship removed and original capabilities preserved — strictly following your instructions.
Models
Hosted models
Every model has its censorship removed while keeping its original capabilities intact.
GLM 5.3 · Ablated
In developmentAblation of the latest GLM flagship is underway. Same promise: unconditional compliance, zero capability tax. Get notified when it ships.
Release progress65%
- Flagship
- Coming soon
Approach
Surgical, not trained
Ablation is a weight-editing technique, not a fine-tune. We locate the direction in the model's activation space that produces refusals — and remove it.
01
Locate
Contrasting activations across harmful and harmless prompts isolates the single direction that mediates refusal.
02
Ablate
Every weight matrix is orthogonalized against that direction, erasing refusal behavior at the source.
03
Validate
Benchmarks confirm zero refusals while reasoning, coding and math scores stay at base-model level.
No dataset. No fine-tuning. No capability tax.
Why Ablatehub
Built for full control
Zero refusals
No safety theatre, no moralizing preambles. The model executes your prompt as written.
Capability preserved
Ablation edits weights surgically — benchmark scores remain at base-model level.
Open weights
Download, run and modify anywhere. Your hardware, your rules.
Unconditional compliance
Instruction-following with no hidden filter overriding your intent.
Local-first
Runs fully offline on your own infrastructure. Nothing leaves your machine.
Research transparency
Method, evaluations and known limitations are documented for every release.
FAQ
Questions, answered
What does “ablated” mean?+
Ablation removes the internal direction a model uses to decide to refuse. After ablation, the model no longer produces refusals — it follows the instructions it is given.
Does removing the safety layer hurt performance?+
No. We edit weights directly instead of fine-tuning, so reasoning, coding and math performance stay at base-model level.
How do I get the models?+
Released models are published as open weights and can be downloaded directly. Links are listed on each model card.
Who is responsible for how the models are used?+
You are. Ablatehub distributes modified open-weight models under their original licenses; operators are solely responsible for using them in compliance with applicable laws.