Skip to main content

Local AI glossary

Plain-language definitions for local AI.

Understand the terms used in on-device AI, private model inference, local documents, browser runtimes, and performance guidance.

Terms A to Z

AI inference
The process of using a trained model to produce an answer, classification, or other output from new input. In CuriousLM, normal chat inference with a downloaded compatible model runs on the user's device.
Read the related guide
Cloud AI
AI where a remote service performs inference. Prompts must cross a network boundary even when the provider applies encryption, access controls, or limited retention.
Read the related guide
Context window
The amount of text and other encoded information a model can consider during one generation. A larger context window can use more memory and take longer to process.
Read the related guide
Embedding
A numerical representation used to compare meaning or similarity. Local document tools can use embeddings to find relevant passages without sending the source document to a hosted model.
Read the related guide
First-token latency
The time between sending a prompt and seeing the first generated token. Model loading, prompt length, device memory, and runtime support can all affect it.
Read the related guide
Local AI
AI that performs inference on the device being used instead of sending each prompt to a hosted model API. Setup, downloads, updates, analytics, and optional online features can still use a network.
Read the related guide
Local RAG
Retrieval-augmented generation where document extraction, indexing, retrieval, and model generation are performed locally. The answer is grounded with selected passages from local source material.
Read the related guide
Model artifact
A downloadable file containing model weights and related data needed by a compatible runtime. Artifact size affects storage use and can influence memory requirements.
Read the related guide
Model weights
The learned numerical parameters of an AI model. A local AI app downloads these parameters so a supported runtime can perform inference on the device.
Read the related guide
On-device AI
AI computation performed by hardware in a phone, computer, or other user-controlled device. The term describes where computation happens, not every data or network behaviour of the application.
Read the related guide
Quantisation
A technique that represents model values with lower precision to reduce storage and memory use. It can make local deployment practical, with possible trade-offs in output quality or compatibility.
Read the related guide
WebGPU
A browser API that gives web applications access to graphics and general-purpose compute capabilities. Supported local AI runtimes can use it to accelerate inference in compatible browsers.
Read the related guide

How definitions are maintained

These definitions describe how CuriousLM uses each term. They are reviewed against primary technical sources and the current product implementation. A definition does not imply that every local AI application has the same privacy or network behaviour.