- Abliteration. Removing a model's refusal direction from its weights so it stops refusing.More
- Active parameters. The parameters a mixture-of-experts model uses for each token, such as 18B of GLM-5.3 Flash's 320B.More
- Anthropic Messages API. The request format Claude Code speaks, at POST /v1/messages.More
- Blended price. Three parts input price and one part output price, divided by four.More
- Chat Completions. The OpenAI request format most tools accept, at POST /chat/completions.More
- Content filter. A check a host runs on prompts or outputs that blocks a request before or after the model.More
- Context window. How many tokens of input and output the model can hold at once.More
- Filter-free host. A host that serves a model with no content filter on top.More
- Fine-tuning. Training a finished model again on new data, which can teach it to stop refusing.More
- GGUF. The single-file model format llama.cpp and Ollama load.More
- Heretic. An open-source tool that abliterates a model and lets an optimizer pick the edit's settings.More
- Jailbreak. A prompt written to talk a normal model past its refusals.More
- KL divergence. How far a changed model's answers drift from the original's; lower means closer.More
- Max output. The most tokens a model returns in one response.More
- Open weights. Model files anyone can download and run.More
- Quantization. Storing weights in fewer bits, such as 4, so a model fits a smaller GPU.More
- Reasoning effort. A setting, low, high or max, for how long a model thinks before answering.More
- Refusal direction. One direction in a model's activations that, when present, makes it refuse.More
- Responses API. OpenAI's newer request format, used by Codex, at POST /responses.More
- Unrestricted AI. A model whose refusal behavior was removed, so it does work mainstream assistants refuse.More