Abliterate LLM weights: plan, modify and evaluate
To abliterate LLM weights, you need a compatible checkpoint and a way to measure changes. This guide outlines a local workflow, with primary implementation resources. To use an existing hosted edition, Uncensured AI provides chat and API access.
Understand what you are changing
Abliteration targets refusal behavior inside a model. It is different from changing a system prompt. Maxime Labonne's tutorial explains a directional method and weight orthogonalization; it is a concrete implementation reference, not a guarantee that the same procedure fits every checkpoint.
Define success before modifying weights: for example, fewer unnecessary refusals on your workload while retaining instruction following. Measure answer quality separately; reducing refusals does not itself demonstrate greater capability.
1. Check the checkpoint, license and hardware
Record the repository, revision, tokenizer and chat template. Confirm access to weights and review license conditions for modification, intended use and distribution. A hosted API key does not give you access to underlying weights.
Check architecture support and memory requirements against your hardware. Keep the original checkpoint intact and reserve space for a separate output, so you can repeat comparisons or restore the starting version.
2. Save a baseline and a held-out evaluation set
Run the original with the intended chat template. Save outputs, settings and timings. Separate examples used to select modification settings from a held-out set for final evaluation. Include ordinary tasks and prompts relevant to the refusal behavior under study.
Define how to score refusals and useful answers. Add workload checks such as code correctness, language consistency, structured output or tool arguments. Apply the same checks to both versions.
3. Choose a tool, configuration and datasets
Heretic's current repository describes several modification objectives and configurations. Older directional-method instructions are not a complete description of today's tool. Use its primary documentation for supported models, setup and commands for your version.
Review datasets and scorers against your objective. Record the tool revision and configuration. Start with a small trial to catch loading, formatting or resource problems before a longer experiment.
4. Modify a versioned copy and compare results
Save candidates separately with versioned configurations. Evaluate held-out examples using the baseline's template and settings. Inspect individual failures: a changed refusal rate can conceal regressions on ordinary requests.
Recheck instruction following, tools, latency and memory in your intended runtime. Retain candidates whose measured tradeoffs fit your objective. If you tune after reviewing final examples, reserve fresh examples for the next comparison.
5. Record the artifact or choose hosted access
Save source and output revisions, tool, configuration, datasets and results with the artifact. Review its license before sharing. These records help someone reproduce the work and understand which conclusions your evaluation supports.
For hosted use, compare the Abliterated editions on Uncensured AI. Chat and the OpenAI-compatible API share a prepaid balance. Hosted access lets you use a listed model; it does not modify your checkpoint or provide downloadable weights.
Models available on Uncensured AI
The list below follows the current catalog. Open a model page for its API ID and current token prices.
- GLM 5.3 Flash Abliterated
glm-5.3-flash-abliterated - GLM 5.3 Abliterated
glm-5.3-abliterated - DeepSeek V4.1 Flash Abliterated
deepseek-v4.1-flash-abliterated - DeepSeek V4 Pro Uncensored
deepseek-v4-pro-uncensored - Qwen 3.8 Max Uncensored
qwen3.8-max-uncensored - GLM 5.3 Uncensored
glm-5.3-uncensored - GLM 5.3 Flash Uncensored
glm-5.3-flash-uncensored - DeepSeek V4.1 Flash Uncensored
deepseek-v4.1-flash-uncensored
Questions and answers
Can I abliterate an LLM through a chat API?
A chat API submits requests to a hosted model. Local modification requires checkpoint access and a compatible tool. Calling an Abliterated edition does not modify your own weights.
How much GPU memory do I need?
Requirements vary by checkpoint and tool. Check current documentation for your architecture, precision and configuration, then verify that your hardware can load and process the model.
Does abliteration preserve the original model's quality?
Do not assume it does. Compare ordinary tasks, instruction following and any tools your application uses against a saved baseline. A lower refusal rate and a higher useful-answer rate are different measurements.
Where are the commands for the actual modification?
Use the linked tutorial and current tool documentation. Commands depend on version, architecture and environment; this guide provides an evaluation workflow rather than a universal installation recipe.