Updated Abliterated LLM API glossary
Short definitions for the terms you meet when you call an abliterated model over an API. Each links to the page that uses it.
- Abliteration
- Removing the one direction in a model's activations that makes it refuse, by editing its weights.
- Source
- Refusal direction
- A single direction in a model's residual stream that mediates refusal, found by Arditi et al. in 2024.
- Source
- Abliterated model
- An open-weight model with its refusal direction removed, so it declines far less often.
- Uncensored model
- A wider label for any model with fewer refusals, by abliteration or by fine-tuning.
- Heretic
- An open-source tool that abliterates models automatically.
- Source
- OpenAI-compatible API
- An API that takes OpenAI's request shapes, so the OpenAI SDKs work after a base URL change.
- Read more
- Chat Completions
- OpenAI's POST /chat/completions format, with a messages array in and choices out.
- Read more
- Messages API
- Anthropic's POST /v1/messages format, used by Claude Code and the Anthropic SDKs.
- Read more
- Responses API
- OpenAI's newer POST /responses format, with input in and output items back.
- Read more
- Base URL
- The address an SDK sends requests to. Changing it is how you switch providers.
- Read more
- Model id
- The string you pass as model, such as kimi-k3.1-pro-uncensored.
- Read more
- Context window
- The most tokens one request can hold, prompt and reply together.
- Read more
- Max output tokens
- The longest reply a model can write in one request.
- Read more
- Reasoning effort
- How much the model thinks before answering, set per request.
- Read more
- reasoning_content
- The field that carries the model's thinking in Chat Completions.
- Read more
- Server-sent events
- The streaming format, with one data line per chunk.
- Read more
- Prompt caching
- Repeated input read from cache and billed at a lower rate.
- Read more
- Per 1M tokens
- The unit API prices use, in dollars per million input or output tokens.
- Read more
- Concurrency limit
- How many requests one account may run at the same time.
- Read more