- 7API
- Abliterated AI
Abliterated AI
An abliterated model is a language model whose refusal behavior has been removed by editing its weights directly, without retraining. The name combines "ablate" and "obliterate". 7API serves abliterated models, including GLM-5.3 Abliterated, through an OpenAI- and Anthropic-compatible API, so you can use one without downloading weights or renting GPUs.
How abliteration works
The technique comes from research showing that refusals in many chat models are driven by a single direction in the model's internal activations (Arditi et al., 2024, "Refusal in Language Models Is Mediated by a Single Direction"). Abliteration uses that finding in three steps:
- Find the refusal direction. Run the model on a set of prompts it refuses and a set it answers, record the activations at each layer, and take the difference of the averages. That difference vector is the refusal direction.
- Remove it from the weights. Orthogonalize the matrices that write into the model's residual stream against that direction, so the model can no longer represent "I should refuse" in that way.
- Check the result. Measure the refusal rate and general benchmarks again. Some builds add a short fine-tune afterward to recover any lost quality.
Because only one direction is removed, the rest of the model (its knowledge, reasoning and tool use) is left largely as it was. The method was popularized for open-weights models in 2024, for example in the Hugging Face write-up "Uncensor any LLM with abliteration".
Abliterated vs. uncensored vs. jailbreaks
| Abliterated | Uncensored (general) | Jailbreak prompt | |
|---|---|---|---|
| What changes | The weights: the refusal direction is removed | Varies: abliteration, fine-tuning or system-level configuration | Only the prompt |
| Works across prompts | Yes, built into the model | Usually | Fragile, breaks with model updates |
| Extra prompt tokens | None | None or a few | Often long, on every request |
| Effect on capability | Small. Some builds recover it with a short fine-tune | Depends on the method | None, but the model may still refuse |
"Uncensored" is the umbrella term. Every abliterated model is uncensored, but not every uncensored model is abliterated. See Uncensored AI API for the full list of both kinds on 7API.
What to expect
- Fewer refusals, not zero. Abliteration removes the main refusal mechanism. Some prompts may still produce a hedge or a disclaimer, and behavior varies between builds.
- Similar capability. Reasoning, coding and long-context work stay close to the base model. Leaderboards don't test abliterated builds, so evaluate on your own tasks.
- No built-in safety layer. The model won't stop you from generating harmful content. You're responsible for how the output is used, and the law applies as usual.
Abliterated models on 7API
| Model | ID | Context | Max output | Input / 1M | Output / 1M | Cached / 1M |
|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 Abliterated | deepseek-v4.1-flash-1017-abliterated | 1M | 66K | $0.15 | $0.15 | $0.05 |
| DeepSeek V4.1 Flash 1017 Abliterated Fast | deepseek-v4.1-flash-1017-abliterated-fast | 1M | 66K | $0.70 | $0.70 | $0.16 |
| GLM 5.3 Abliterated | glm-5.3-abliterated | 1M | 130K | $0.80 | $0.80 | $0.25 |
All of them have a 1M-token context window and work on Chat Completions, Responses and Anthropic Messages, so they run in Claude Code, OpenCode, Codex, Cursor and other OpenAI-compatible tools.
Call an abliterated model
curl https://7api.st/v1/chat/completions \
-H "Authorization: Bearer $SEVENAPI_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "glm-5.3-abliterated", "messages": [{"role": "user", "content": "…"}]}'In Claude Code, run /model claude-glm-5.3-abliterated. The abliterated GLM models always think, at high effort or above. Send "reasoning_effort": "xhigh" or "max" for harder tasks. See Reasoning effort.
Why use an API instead of running it locally
Abliterated weights for small models can run on a desktop GPU. Frontier-size models such as GLM-5.3 need a multi-GPU server to run at useful speed with a long context. Through an API you pay per token, with cached input billed at a lower rate, and switch between abliterated, uncensored and standard models by changing one string.
Common uses
- Security research and red-teaming: generating test cases and attack scenarios that standard models refuse.
- Refusal research: comparing how models with and without refusal behavior handle the same prompts.
- Fiction and creative writing with mature or dark themes.
- Research on sensitive topics such as medicine, drugs and extremism, where standard models over-refuse.
Is abliterated AI legal?
Using a model with fewer refusals is legal in most places. What you generate with it is still subject to the law where you live. Abliteration changes what the model will answer, not what is allowed.
What does "abliterated" mean?
It means the model's refusal direction has been identified and removed from its weights. The word blends "ablation", removing part of a network to study its effect, with "obliterate".
Is GLM-5.3 Abliterated the same as GLM-5.3 Uncensored?
No. Both refuse less than standard GLM-5.3. GLM-5.3 Abliterated is the weight-edited build with the largest output budget (130,000 tokens). GLM-5.3 Uncensored costs much less and is the better choice when you need volume. See the comparison.
Do abliterated models score lower on benchmarks?
Usually slightly, by a few points on most suites, and some builds recover that with a short fine-tune afterward. Independent leaderboards don't list abliterated builds, so the benchmarks page shows the base models only.
