term
Quantization
Reducing the numeric precision of model weights (e.g. to 4-bit) to use less memory. Can cost a little quality.
Norwegian reference for local AI — registers, plans and honest tests.
Dictionary
Short explanations of words, tools and formats you meet when you build local AI.
Last checked:
Each card has a type, a short explanation and a link. Tools and models show a license from an official source where we have verified it. See also kilder og lisenser.
file format
GGUF
File format used with llama.cpp to pack (often quantized) weights in one file that is easy to load locally.
term
Token
Smallest text piece the model reads or writes. A Norwegian word can be one or more tokens.
term
Kontekstvindu
How many tokens the model can hold at once. A larger window needs more memory (VRAM/RAM).
term
Inferens
Running a trained model to generate answers — as opposed to training.
term
VRAM
Memory on the graphics card. Local LLMs should preferably fit here; otherwise it gets slower via system RAM.
term
Unified memory
Shared memory for CPU and GPU (Mac, Jetson). OS and model share the same budget — leave a margin.
term
Harness
Rammeverk eller agent-skall rundt en model: tool, minne og grensesnitt — ikke selve vektene.
term
Runtime
The software that loads and runs the model (llama.cpp, Ollama, MLX, vLLM …).
tool
Ollama
CLI and local API to fetch and run models, often via GGUF. LM Studio is the usual GUI alternative.
tool
LM Studio
Desktop app with a GUI to search, download and run local models (GGUF/MLX etc.) with a local server.
tool
llama.cpp
C/C++ runtime for local LLMs with broad GGUF support on CPU and GPU.
term
LoRA
Low-Rank Adaptation: light fine-tuning via small adapter matrices instead of the whole model.
term
RAG
Retrieval-Augmented Generation: hent egne dokumentbiter og gi dem til modelen som kontekst.
term
Embedding
Number vector that represents text meaning. Used for search and RAG.
term
TOPS
Tera Operations Per Second — a marketing figure for accelerators. Not the same as tokens/s in practice.
hardware
Jetson
NVIDIA’s on-board computers with a GPU for edge/AI. Unified memory and low power use.
tool
MLX
Apple’s machine-learning framework for Apple Silicon (unified memory), with mlx-lm for language models.
tool
vLLM
Inference engine for high throughput, typically on Linux with an NVIDIA GPU.
model
Whisper
OpenAI’s speech-to-text family. Runs locally via faster-whisper, whisper.cpp and others.
tool
faster-whisper
Fast Whisper implementation (CTranslate2) for local speech-to-text on CPU or GPU.
term
Agent
System som bruker en model med tool og flere steg — mer enn én chat-respons.
term
Finjustering
Training further on a finished model with your own data. Heavier than inference; LoRA is a light variant.
term
Checkpoint
En lagret versjon av modelvekter. Alltid sjekk LICENSE for akkurat den filen.
Frequently asked questions
How many entries are in the dictionary?
The seed has 22 entries. Each is a short definition, not a textbook.
Is this the same as the license or model register?
No. The dictionary explains terms (GGUF, VRAM, Ollama …). Licenses and models have their own registers.
Can I suggest a new word?
The suggestion channel is coming. See Suggestions — no live submission until privacy is approved.
Why do some entries cross-link to other words?
So you can jump quickly between terms you meet together when building local AI.