AbliterantAI research
We build unrestricted open-weight models
We're an AI research lab. We measure each model against the one it came from and publish the results.
What we work on
Removing refusals
We take learned refusals out of a model's weights and keep the rest of what it can do. The edit is called abliteration.
Inside RANAMeasuring the cost
We test every model against the one it came from and against another abliteration of it.
Our researchPublishing the evidence
The weights, model cards, and raw results are public, so anyone can check our numbers.
Weights on Hugging Face
Models
All modelsRANA on Qwen3.5-9B
A 9B vision-language model with a thinking mode. It gives more useful answers than a public abliteration of it and loses about a quarter as much capability.
RANA on Qwen3.8-27B
A 27B vision-language model with a thinking mode. It answers more of the test requests than orca, another abliteration of it, and stays closer to the base model.
Research
All research notesAbliterant Chat
Abliterant Chat answers with abliterated models, which are edited to refuse less: GLM-5.3 and, on Pro, Kimi K3.1 Pro. It saves your conversations encrypted, with keys only you hold, so we can't read them.
Free
$0
Up to 10 chat messages per UTC day.
Pro
$20 USD a month
Up to 200 chat messages per UTC day.
Figure 1. A conversation, as our server stores it
- title
- Sealed, 285 bytes
- your message
- Sealed, 285 bytes
- reply
- Sealed, 1,757 bytes
- conversation key
- Sealed, wrapped by your account key
- created
- messages
- 2: one from you, one complete reply
- pinned, archived
- no, no
- Sealed: only your keys open it
- Visible to us
A demo conversation, “What does a Unix pipe do?”, as our account API returns it to the browser, shortened. Red marks what's sealed with AES-256-GCM.
Source: our account API's record of the conversation on Nov 1, 2026, since deleted.
Join the community
Talk about the models and the research with the lab and with other people who run them.