Glossary · Category
Additional Safety, Evaluation & Model Behavior Terms
10 plain-English definitions.
- Distillation restriction clause
- A license term (present in some model releases) that restricts using a model's outputs to train or improve competing models.
- Gated/restricted model release
- A release pattern where model weights require accepting usage terms or an approval process before download, as opposed to fully unrestricted open access.
- Model lineage
- The documented chain of base models, fine-tunes, and merges that a given model was derived from, relevant for licensing, attribution, and capability inheritance.
- Model weights (open vs closed)
- The trained numerical parameters of a model; "open-weight" models release these files for download and local use, while "closed" models only expose access via a hosted API.
- Open-weight frontier model
- A model whose weights are publicly released while still competing with closed frontier models on capability benchmarks, a category that has grown significantly (e.g., DeepSeek, Qwen, Llama, Kimi).
- Open-weight license
- The specific legal terms under which a model's weights are released (e.g., Apache 2.0, MIT, or custom community licenses with usage restrictions).
- Permissive license (model)
- A license (e.g., Apache 2.0, MIT) placing minimal restrictions on commercial use, modification, and redistribution of a model's weights.
- Reasoning effort parameter
- An API parameter that lets a caller control how much internal reasoning/thinking compute a model spends on a request, trading off latency and cost against answer quality.
- Thinking mode / hybrid reasoning model
- A model that can toggle between a fast direct-answer mode and a slower extended chain-of-thought "thinking" mode, often via a prompt flag or dedicated model variant.
- Weight diff / delta weights
- A distribution method that releases only the numerical difference between a licensed base model's weights and a fine-tuned version, requiring users to already hold the base weights.