Abliterated models
Abliterated models for high-risk work.
Use abliterated models with capability evaluations, zero data retention by default, and Policy Gateway for rules you control on model calls.
What is an abliterated model?
An abliterated model is a language model edited to reduce learned refusal behavior. Abliteration gives you hosted access to these models for security research, finance, healthcare, and other work where provider refusals can interrupt legitimate tasks.
Capability retention is part of the evaluation.
We measure refusal reduction and task performance separately. Our evaluations cover reasoning, coding, and agent tasks, so we can assess whether a model still does useful work after its refusal behavior changes. We publish results for the specific model being served.
Published capability results.
Abliterated Large v2 · GLM 5.3 · FP8
- Terminal-Bench 4.0
- 41.8%Resolution rate
- CyberGym
- 84.5%Pass@1 on 1,507 tasks
- ExploitGym
- 105 / 869Tasks completed in 2 hours, TPS-normalized
Zero data retention by default.
Your model API prompts and responses are processed in memory and discarded after the request. They are never used to train models. We retain operational metadata for billing and reliability, not the content of your model calls.
Data handling and retentionSet your own rules with Policy Gateway.
Write rules for what your application allows. Policy Gateway can monitor matches or block matching requests and responses, with content-free decision records. Access is included with every plan, including pay as you go.
Explore Policy GatewayKeep your OpenAI or Anthropic SDK.
Connect through compatible APIs using your existing client. Change the base URL, API key, and model selection, then check the compatibility guide for the endpoints and request fields your application uses.
API compatibilityTry an abliterated model on your own work.
Choose a model, bring a task with a known result, and compare the answers, tool use, and cost. Review the published evaluations alongside the requirements of your application.
Try AbliterationFAQ
Frequently asked questions.
Does abliteration affect model capabilities?
It can, which is why we evaluate capability retention alongside refusal reduction. We publish model-specific results for reasoning, coding, and agent tasks. Those results describe the tested model and settings; evaluate your own tasks when choosing a model.
Which abliterated model should I use?
Compare the current models by input type, context window, supported features, and price. Our model documentation lists the available model IDs and their capabilities. Use the exact model you plan to deploy when testing your workload.
Are my model API prompts and responses stored or used for training?
No. Core model API prompts and responses are processed transiently and discarded after the request, and are never used to train models. Operational metadata such as token counts, timestamps, and status codes is retained for billing and reliability.
Does zero data retention apply to every platform feature?
Zero data retention applies to core model inference. Optional web search and fetch send queries or URLs to external providers whose retention policies apply. The training-data generator stores datasets for you to review and export. See the data handling policy for the scope of each feature.
Can I set my own rules on model calls?
Yes. Policy Gateway evaluates requests and responses against policies you define. Start in Monitoring to review matches, then use Enforcing to block matching content. Policy Gateway stores policy configuration and content-free decision metadata, not your prompts or responses.
Can I use the OpenAI or Anthropic SDK?
Yes. Use our compatible APIs with your existing SDK, an Abliteration API key, and an available model ID. Check the compatibility matrix for supported endpoints and request parameters.