What makes an abliterated model the best?
The best abliterated model for you is the one with the fewest refusals and the smallest quality loss, on a base model suited to your task, at a price that works for your volume. Those three things pull in different directions, so “best” depends on what you are building.
Abliteration removes a learned refusal direction from a model’s weights. If you want the mechanics, the guide to abliterated models covers them. What matters here is that the edit is blunt. It touches every layer, and two people abliterating the same base model can get noticeably different results depending on which layers and directions they pick and whether they fine-tune afterward.
So the name tells you less than you’d hope. “GLM-5.3 Abliterated” on one provider and “GLM 5.3 Abliterated” on another may be different builds. Judge each one on its own.
How to judge an abliterated LLM
Judge an abliterated LLM on five things, roughly in this order:
- Base model. Abliteration can’t make a model smarter. A strong base model with a careful abliteration beats a weak base model with a perfect one. Check which release it’s built on, and its size.
- Refusal rate. Count how often it still declines prompts the base model refuses. Abliteration cuts refusals a lot, but leftover refusals on certain phrasings are common.
- Distance from the base model. On harmless prompts, a good abliteration should answer much like the original. Tools measure this with KL divergence, which compares the two models’ next-token probabilities. A low number means the edit changed little beyond refusals.
- Healing. Some builders run a short fine-tune after abliteration, often DPO, to recover lost quality. A build that says it was healed is usually a safer bet than a raw one.
- Serving details. Context window, quantization and price. A heavily quantized build can be cheaper and worse, and a provider may not say which quantization it serves.
The open-source tool Heretic, by Philipp Emanuel Weidmann, makes the second and third points concrete. It searches for abliteration settings that minimize both the refusal rate and the KL divergence from the original model at the same time. When a model card reports those two numbers, you can compare builds of the same base model directly. Hosted APIs rarely publish them, so in practice you measure refusals and quality yourself.
Which hosted abliterated models does the directory list?
This directory lists 7 abliterated models across Refuseless, Unaligned, UncensoredLLM and NanoGPT. Here they are, cheapest first on a typical roleplay workload, with published prices per 1 million tokens:
| Provider and model | Input / output per 1M | Context | Per 1,000 roleplay turns | Per 1,000 long replies |
|---|---|---|---|---|
| RefuselessDeepSeek V4.1 Flash 1017 Abliterated | $0.12 / $0.45 | 1M | $0.855 | $1.08 |
| RefuselessGLM 5.3 Flash Abliterated | $0.12 / $0.50 | 1M | $0.87 | $1.18 |
| RefuselessGLM 5.3 Abliterated | $1.00 / $2.00 | 1M | $6.60 | $5.50 |
| NanoGPTAbliterated Large Flash | $1.00 / $2.00 | Not listed | $6.60 | $5.50 |
| UnalignedGLM-5.3 Abliterated | $2.00 / $3.20 | 1M | $12.96 | $9.40 |
| UncensoredLLMGLM-5.3 Abliterated | $2.183 / $4.365 | Not listed | $14.405 | $12.004 |
| NanoGPTAbliterated Model Large V2 | $3.00 / $5.00 | Not listed | $19.50 | $14.50 |
Three things stand out. First, the spread is wide: the most expensive model costs about 23 times the cheapest per roleplay turn. Second, most of the list is built on GLM 5.3 or DeepSeek V4.1. Third, Abliterated Large Flash and Abliterated Model Large V2 on NanoGPT don’t name their base model, so ask the provider or check its model page before you rely on them. NanoGPT also says its abliterated models are offered for authorized and defensive work under its terms, so read those before you build on one.
Refuseless and Unaligned also publish cached input rates for these models, from $0.025 per 1M tokens. For chat apps that re-send long histories, that can cut the bill well below the table above. The pricing guide shows how to work it out.
Flash or full size: which abliterated model size do you need?
Start with a Flash model and move up only if it fails on your prompts. Flash models are the smaller, faster versions of a provider’s flagship, and they cost a fraction as much.
On Refuseless, GLM 5.3 Abliterated costs $1.00 per 1M input tokens and GLM 5.3 Flash Abliterated costs $0.12. That’s about 8 times more on input and 4 times more on output. For a 50,000-turn roleplay month, it’s the difference between $43.50 and $330.00.
What you get for the extra money is mostly better reasoning, steadier instruction following over long chats and fewer odd turns in hard prompts. Abliteration tends to hurt those same skills a little, so a larger base model has more room to absorb the loss. For short replies, casual chat and high volume, a Flash build is usually enough. For long stories, complex character cards with many rules, or anything that needs multi-step reasoning, test a full-size model too.
Count regenerations when you compare. Every reply a user throws away is paid for, so compare the cost per reply your users keep, not the cost per request.
Best abliterated model by use case
Based on published prices and model size, here is where to start for common jobs. These are starting points for your own tests, not quality rankings.
| Use case | Start with | Why |
|---|---|---|
| High-volume chat or roleplay | DeepSeek V4.1 Flash 1017 Abliterated on Refuseless | Lowest roleplay cost in the list. Context window: 1M. |
| Cheap GLM-based roleplay | GLM 5.3 Flash Abliterated on Refuseless | Same price tier as the cheapest option, on a different base model. Worth testing side by side. |
| Long stories and complex character cards | GLM 5.3 Abliterated on Refuseless | Full-size model, the lowest-priced full GLM 5.3 abliterated build in the list. |
| Text, image and video on one key | GLM-5.3 Abliterated on UncensoredLLM | Same balance also pays for image and video models. |
| Crypto-only, no email signup | GLM-5.3 Abliterated on Unaligned | Account without an email. Also publishes a cached input rate. |
For card payments, Refuseless, UncensoredLLM and NanoGPT accept cards. Unaligned takes crypto only.
Whichever you pick, the provider’s terms and the law still apply. Abliteration only changes the model’s tendency to refuse. Sexual content involving minors, and intimate images of real people made without consent, are illegal, and no model or provider makes them acceptable.
How to test abliterated models yourself
You can compare abliterated LLMs in an afternoon with a fixed prompt set and a spreadsheet. All the providers above are prepaid and per-token, so a test run costs cents.
- Build a prompt set. Take 20 to 50 real prompts from your app. Include ones a standard model refuses and ordinary ones that it handles fine.
- Fix the settings. Use the same system prompt, temperature and max tokens for every model. Change only the model ID. All of these providers offer OpenAI-compatible endpoints, so that’s one string, as the OpenAI-compatible API guide shows.
- Count refusals. Mark each reply that declines, lectures, or quietly dodges the request. Soft refusals, where the model changes the subject, count too.
- Judge quality on the ordinary prompts. This is your stand-in for KL divergence. If a model stops refusing but answers normal prompts worse than a non-abliterated model of the same family, the abliteration cost too much.
- Check long conversations. Run a few 30 to 50 turn chats. Some abliterated builds drift, repeat themselves or lose the character over long contexts.
- Log tokens. Read
prompt_tokensandcompletion_tokensfrom each response and price the run with the formula from the table above. - Pick a main model and a fallback. Providers retire and rename models, so keep a second abliterated model, ideally on another provider, ready to swap in.
Re-run the same set when a provider changes a model. A new build under the same name can behave differently.
Frequently asked questions
What is the best abliterated model?
There isn’t one best abliterated model for every job. A good one refuses rarely and loses little quality compared with its base model. On price, DeepSeek V4.1 Flash 1017 Abliterated on Refuseless is the cheapest hosted abliterated LLM in this directory, and full-size GLM 5.3 abliterated builds are the next step up for harder prompts.
What is the cheapest abliterated LLM API?
DeepSeek V4.1 Flash 1017 Abliterated on Refuseless is the cheapest abliterated model in this directory, at $0.12 per 1M input tokens and $0.45 per 1M output tokens. A 50,000-turn roleplay month at 6,000 input and 300 output tokens per turn costs about $42.75 on it before caching.
How do you measure the quality of an abliterated model?
Measure two things: how often it still refuses prompts the base model declines, and how closely it matches the base model on harmless prompts. Abliteration tools such as Heretic express the second as KL divergence. With a hosted API, run the same prompt set on each model and compare.
Are abliterated LLM models better than uncensored fine-tunes?
Not as a rule. Abliteration keeps the base model’s knowledge and style while removing refusals, and fine-tunes can change the writing style, sometimes for the better in fiction. In this directory the abliterated builds also cost more than uncensored builds of the same base model, so test both.
Which providers host abliterated models?
In this directory, Refuseless, Unaligned, UncensoredLLM and NanoGPT host abliterated models. All of them sell prepaid, per-token access, and all of them offer OpenAI-compatible endpoints.