Guides · Updated
Local AI glossary
40 terms you'll meet when you start running AI on your own computer, explained in plain English. Each one links to the guide that covers it in more depth.
A
Agent (tool calling)
A setup where the model doesn't just reply but takes actions: reading files, running commands or calling other programs ("tools"), then using the results. Only some models are trained for tool use; Ollama shows a tools capability on each model's library page.
Learn more: Running local AI agents safely on your PC · Set up a free local coding assistant (Continue, VS Code, Cursor) · Ollama models explained
API (OpenAI-compatible)
The way other programs talk to your model. Ollama serves one at http://localhost:11434 and LM Studio on port 1234. Both offer an OpenAI-compatible API, so many apps written for OpenAI's cloud service can use your local model just by changing the address.
Learn more: LM Studio vs Ollama
C
Checkpoint
A saved copy of a model's weights. In image generation it means the main model file you load before generating, such as an SDXL checkpoint. In ComfyUI these go in the models/checkpoints folder.
Learn more: Generating images locally
Context length (context window)
How much text the model can take into account at once: your messages, its replies and any documents, measured in tokens. When a chat goes over the limit, the oldest content is dropped, often without any warning. The app's setting can be much smaller than the model's maximum, and a longer context uses more memory.
Learn more: When it doesn't work · VRAM and quantisation explained
CUDA
NVIDIA's platform for running general computing work on its graphics cards. Most AI tools support it first, which is why NVIDIA cards are usually the easiest option. For Ollama, what you mainly need is an up-to-date NVIDIA driver.
Learn more: How to install Ollama on Windows (2026) · When it doesn't work
D
Diffusion model
The kind of model behind most image generators, such as Stable Diffusion. It starts from random noise and refines it step by step into an image that matches your prompt. Image models run in apps like ComfyUI, not in Ollama.
Learn more: Generating images locally
Docker
A tool that runs an app in a self-contained "container" with everything it needs bundled in. Open WebUI's recommended install uses Docker, so you don't have to set up Python yourself. A container reaches Ollama on the same PC through host.docker.internal rather than localhost.
Learn more: Running a local AI chat with a nice interface (Open WebUI and alternatives) · When it doesn't work
E
Embeddings
Lists of numbers (vectors) that capture the meaning of a piece of text, so texts with similar meanings get similar numbers. Embedding models such as nomic-embed-text create them, and they power the search step in RAG. Embedding models aren't chat models, and vectors from different embedding models can't be mixed.
Learn more: Chat with your documents locally
F
Fine-tuning
Training an existing model further on extra data so it gets better at a particular job or style. Coding versions of chat models and many community models are fine-tunes. You run a fine-tuned model the same way as the original.
Learn more: Fine-tuning a small model with LoRA · Ollama models explained
G
GGUF
A single-file model format that came from llama.cpp and is used by Ollama, LM Studio and many other apps on every platform. A GGUF file holds the model's weights, usually quantised, plus the details needed to run it. Ollama can run GGUF models from Hugging Face directly.
Learn more: Ollama models explained · Local AI on a Mac with Apple Silicon
H
Hallucination
When a model states something false with confidence, such as an invented fact, quote or source. Every language model does it, and smaller models tend to do it more. Giving the model your own documents (RAG) and telling it to answer only from them helps, but always check anything important.
Learn more: Chat with your documents locally
Hugging Face
The main website where AI models are published and shared, a bit like GitHub for models. Most open models appear there, often in several formats and quantisations. LM Studio has a Hugging Face browser built in, and Ollama can pull GGUF models from it. Prefer the original publisher or well-known uploaders.
Learn more: Ollama models explained · LM Studio vs Ollama
K
KV cache
The model's working memory for the current conversation. It stores information about every token so far so the model doesn't have to recompute it, and it grows with context length. That's why long chats and documents need memory on top of the model itself. Ollama can quantise it to save space.
Learn more: VRAM and quantisation explained
L
llama.cpp
An open-source engine for running language models efficiently on ordinary hardware: CPUs, NVIDIA, AMD and Intel GPUs, and Apple Silicon. It's where the GGUF format comes from, and many apps, including LM Studio and Ollama, build on it.
Learn more: LM Studio vs Ollama
LM Studio
A desktop app for finding, downloading and chatting with local models, with a built-in model browser and settings you can see and adjust. It's free for personal and work use; the app itself isn't open source.
Learn more: LM Studio vs Ollama
localhost
Your own computer's network address (also written 127.0.0.1). A service listening on localhost can only be reached from the same machine. Ollama listens on localhost:11434 by default. It has no login of its own, so keep it that way unless you've protected it with a VPN or an authenticating proxy.
Learn more: Local AI for privacy
LoRA
Short for low-rank adaptation: a small add-on file that adjusts a model without retraining all of it. In image generation, LoRAs teach a base model a style, character or subject and are loaded alongside a checkpoint. A LoRA is made for one base model (for example SDXL) and generally won't work with another. Language models can be fine-tuned with LoRAs too.
Learn more: Fine-tuning a small model with LoRA · Generating images locally
M
Memory bandwidth
How fast data can be read from memory, measured in gigabytes per second. Writing each token means reading almost all of the model's weights, so bandwidth sets a rough ceiling on speed: tokens per second ≈ bandwidth ÷ model size. Graphics card memory is many times faster than system RAM.
Learn more: VRAM and quantisation explained
Mixture of experts (MoE)
A model split into many "experts", of which only a few are used for each token. A name like 30b-a3b means 30 billion parameters in total, with about 3 billion active. You still need memory for the whole model, but it runs at roughly the speed of a much smaller one and copes better when split between GPU and system RAM. Models that use all their parameters for every token are called dense.
Learn more: VRAM and quantisation explained · Ollama models explained
MLX
Apple's open-source machine-learning framework for Apple Silicon. MLX versions of models often run a little faster on a Mac than GGUF ones. LM Studio supports MLX models, and the mlx-lm tool runs them from the command line.
Learn more: Local AI on a Mac with Apple Silicon
Model tag
The part after the colon in an Ollama model name, as in qwen3.5:9b. It usually picks the size, and sometimes a variant or quantisation. With no tag, Ollama uses latest, which points to the model's default size.
Learn more: Ollama models explained
N
NPU
Neural processing unit: a low-power AI chip built into many recent laptops and "AI PCs". Popular local chat tools like Ollama mostly don't use it, so it isn't worth paying extra for if running chat models is your goal.
Learn more: Local AI on a budget PC (no big GPU)
O
Offloading (GPU/CPU split)
Splitting a model between the graphics card and system RAM when it doesn't fit in VRAM. It still works, but it's much slower, because system RAM is slower than VRAM. In Ollama, ollama ps shows the split; LM Studio calls the setting GPU offload.
Learn more: When it doesn't work · VRAM and quantisation explained
Ollama
A free, open-source program for Windows, Mac and Linux that downloads and runs local models with simple commands like ollama run. It runs quietly in the background, so chat apps, coding assistants and scripts can connect to it.
Learn more: How to install Ollama on Windows (2026) · LM Studio vs Ollama
Open WebUI
A self-hosted chat interface that looks and works much like ChatGPT. You use it in your browser, but it's served from your own PC. It connects to Ollama and adds chat history, document upload and user accounts.
Learn more: Running a local AI chat with a nice interface (Open WebUI and alternatives)
Open-weight model
A model whose trained weights are published so you can download and run it yourself, such as models from the Llama, Qwen, Gemma, Mistral and gpt-oss families. Open weights don't always mean open source: licences vary, and some restrict commercial use. Check the licence on the model card.
Learn more: Ollama models explained
P
Parameters
The numbers inside a model that were learned during training, also called weights. Model size is counted in them: "9B" means about 9 billion. More parameters usually means a more capable model, but also more memory and slower replies.
Learn more: VRAM and quantisation explained · Ollama models explained
Q
Q4_K_M (quantisation labels)
A quantisation label from llama.cpp's GGUF format. The number is roughly the bits per weight (Q4 is about 4-bit, Q8 about 8-bit), K means the "k-quant" method, and S, M or L means a small, medium or large share of weights kept at higher precision. Q4_K_M is Ollama's usual default: a good balance of size and quality.
Learn more: VRAM and quantisation explained
Quantisation
Storing a model's weights with fewer bits so it takes less memory, at the cost of a little quality. At full 16-bit precision an 8B model needs about 16 GB; at 4-bit it shrinks to roughly 5 GB. A bigger model at 4-bit generally beats a smaller one at 8-bit.
Learn more: VRAM and quantisation explained
R
RAG (retrieval-augmented generation)
A way for a model to answer from your own files. Documents are split into chunks and turned into embeddings; when you ask a question, the most relevant chunks are found and added to the prompt, so the model can answer from them. Done locally, your documents never leave your machine.
Learn more: Chat with your documents locally · Local AI for a small business
Reasoning model
A model trained to work through a problem step by step ("thinking") before it answers. That helps with maths, logic and planning, but makes simple chat slower. DeepSeek-R1 and gpt-oss are examples. Some models and apps let you turn thinking off for quick questions.
Learn more: Ollama models explained
ROCm
AMD's platform for running AI work on Radeon graphics cards, its counterpart to NVIDIA's CUDA. It's the fastest way to use a supported AMD card with Ollama, but only certain cards are supported, and the list differs between Windows and Linux.
Learn more: Running local AI on AMD, Intel Arc and mini PCs
S
Safetensors
A model file format (.safetensors) that holds only the weights, so loading it doesn't run hidden code. Older .ckpt and .pt files can contain code that runs when they're loaded, so prefer safetensors, especially from sources you don't fully trust.
Learn more: Generating images locally
Speech-to-text and text-to-speech
Speech-to-text turns recordings into written text; Whisper is the open model most local transcription tools are built on. Text-to-speech does the reverse and reads text aloud, with tools such as Piper and Kokoro. Both use separate apps from Ollama, and many run well on a CPU.
Learn more: Local speech-to-text and text-to-speech on your PC · Build a private voice assistant on your own PC
T
Tokens
The chunks of text a model reads and writes: often a whole short word, sometimes part of a longer word or a punctuation mark. In English a token is, on average, a bit less than a word. Context length and speed are both measured in tokens.
Learn more: VRAM and quantisation explained
Tokens per second
How fast a model writes its reply. It depends mostly on memory bandwidth and model size, and drops sharply when a model doesn't fit fully in VRAM. Speeds shown on this site are estimates, not benchmarks.
Learn more: VRAM and quantisation explained · GPU model picker
U
Unified memory
Memory shared by the CPU and GPU, as on Apple Silicon Macs and some AMD mini PCs. The GPU can use most of it, so a machine with lots of unified memory can load models bigger than most graphics cards can hold, though usually with less bandwidth than a high-end card.
Learn more: Local AI on a Mac with Apple Silicon · Running local AI on AMD, Intel Arc and mini PCs · Choosing a mini PC for local AI
V
Vision model
A model that reads images as well as text: it can describe photos, read screenshots or pull text out of a picture. Vision models read images; they don't create them. Ollama shows a vision capability on models that support it.
Learn more: Ollama models explained
VRAM
Video RAM: your graphics card's own memory. The model, plus memory for the conversation, needs to fit in it to run at full speed, which makes it the biggest factor in which models you can run. On Windows, Task Manager → Performance → GPU shows it as Dedicated GPU memory.
Learn more: VRAM and quantisation explained · GPU model picker
Vulkan
A cross-vendor graphics standard that Ollama, LM Studio and llama.cpp can use to run models on many AMD, Intel and other GPUs. It covers more cards than ROCm, and it's how Intel Arc cards run in Ollama.
Learn more: Running local AI on AMD, Intel Arc and mini PCs
Get the Local AI weekly
New models worth running, what fits which GPU, and setup fixes. One short email a week, free. Unsubscribe any time.
Local-LLM App Starter
A tested starter for building your own app on Ollama: chat, RAG and a GPU-fit checker. Skip the setup hours.
kits.lilaforge.com → VibeEngineBest local LLM tools, ranked
Ollama, LM Studio, Open WebUI and llama.cpp scored from public evidence, with top picks.
vibeengine.lilaforge.com →