Concept

Open-source AI concepts

Plain definitions of the terms that come up in open-source AI projects, each linked to the projects in our catalogue that implement them.

Reading paths

The concepts in the order they build on each other. Start at the first and each page hands you to the next.

What is a Copyleft license?A copyleft licence grants you broad permission to use, modify and redistribute a program, and attaches one condition in return: if you convey a modified version, your recipients must get the same freedom over your changes. GPL, AGPL, LGPL and MPL are that same idea with different scopes, and the whole practical question is where the boundary of your version sits.What is a Diffusion model?A diffusion model is a neural network trained to undo noise: it learns to predict what was added to an image at each step of a fixed noising process, so generation starts from random noise and is repeatedly denoised into a picture. Latent diffusion, the variant behind Stable Diffusion, runs that loop inside a compressed space with the text prompt as conditioning, which is why it fits on a consumer GPU and why the tooling around it is mostly about samplers, schedulers and control.What is a Multi-agent system?A multi-agent system is a setup where several language model agents take different roles and hand work to each other instead of one model answering alone. Open-source implementations range from a fixed pipeline that turns one prompt into a repository to runtime orchestrators and workspace managers built for fleets of coding agents.What is a Permissive license?A permissive licence, such as the MIT licence or Apache-2.0, grants you the right to use, modify, merge, publish and sell a program with no condition attached beyond attribution, so a permissive dependency can be absorbed into a closed product and never returned. That is the trade, and the differences worth knowing inside the family are the patent grant and the notice requirements rather than the prose.What is a Source-available license?A source-available license publishes the source code so you can read it, run it and often change it, while withholding one of the freedoms that define open source, usually the right to use it commercially as a service. That is why the category exists, and it is why adopting one is a business decision rather than a technical one.What is AGPL-3.0?AGPL-3.0 is the GNU Affero General Public License, and it is the GPL copyleft licence with one extra clause: instead of triggering only when you distribute the software, it also triggers when you let users interact with a modified version over a network, which is what running a service is. That single sentence is why a company's legal team reads the licence before a deployment rather than after it.What is AI coding agent?An AI coding agent is a program that uses a large language model to inspect a codebase, edit files and run commands toward a goal you state in natural language. The term also appears as coding agent or agentic coding.What is an AI agent?An AI agent is a program that uses a large language model (LLM) to decide on actions, call tools, observe the results and repeat until a task is finished. This page explains the mechanism, when it is worth using, and how the term appears in open-source projects.What is an Attention mechanism?An attention mechanism is the part of a network that decides which parts of an input matter to which other parts, by scoring every pair of positions and mixing the values according to those scores. Self-attention is the version where that happens inside one sequence, and it is the operation whose cost, and whose kernels, dominate the way modern models are implemented.What is an Open-weight model?An open-weight model is one whose trained parameters are published for download, so anyone can run, inspect or adapt it without asking the developer for an API key. It describes the weights and nothing else, not the training data, not the training code and not necessarily the licence terms, which is why the term gets used loosely and why reading the licence is still the first task.What is Chunking in RAG?Chunking is the step that cuts documents into the pieces a retrieval system indexes and a language model later reads, and it is the step most systems perform once, with a default size, and never revisit. It matters because a chunk cut in the wrong place does not fail: it embeds successfully, retrieves successfully, and hands the model half an idea.What is Computer use?Computer use is the practice of letting an AI model drive a real browser or desktop session, clicking, typing and reading the screen on your behalf. A browser agent applies the same idea to web pages only, usually through a browser automation layer.What is Context window?A context window, also called context length, is the maximum number of tokens a language model can attend to in one request, covering the system prompt, conversation history, tool output and the reply. Tokens beyond that limit are truncated or the request fails, so the window sets a hard ceiling on how much an agent can see at once.What is CUDA?CUDA is NVIDIA's parallel programming model for their own GPUs: you write a kernel, launch thousands of threads that run the same instructions over different data, and take responsibility for where every byte lives in a memory hierarchy where bandwidth, not capacity, is usually the limit. Most people meet it through a framework that calls kernels on their behalf, and the real questions start where the framework has no kernel for what you need.What is Embeddings?Embeddings are dense numeric vectors that represent text, images or other content so that similar meanings land close together in vector space. A text embedding model turns a piece of content into such a vector, which is why search, clustering and recommendation systems depend on them.What is Fine-tuning?Fine-tuning is continued training of an already pretrained model on a smaller, task-specific dataset, so that the model changes its behaviour, output format or domain fit instead of being trained from scratch. The term covers everything from updating every weight in the model to training a small set of added adapter parameters, and those variants differ enormously in what they cost.What is FlashAttention?FlashAttention is an attention kernel that computes exactly the same result as the textbook implementation while never writing the full attention matrix to memory, by tiling the computation and keeping running sums on chip. In open source it arrives as CUDA kernels that have to be matched to your Python, CUDA and PyTorch combination, and that build step is where most of the friction lives.What is Function calling?Function calling is the mechanism by which a language model asks your application to run code on its behalf: you describe the available functions with names and parameter schemas, the model replies with a structured request naming one of them and its arguments, your code executes it, and the result goes back into the conversation. The model proposes the call; your program decides whether to make it and does the running.What is GGUF?GGUF is a binary file format that stores a machine learning model's weights, metadata and tokenizer information in one file so inference software can load it without a separate configuration step. It is the format used by llama.cpp and by many other local inference tools.What is GraphRAG?GraphRAG replaces similarity search over embedded chunks with a graph: a model extracts the entities in your documents and the relations between them, and a question is answered by walking those relations instead of by finding text that looks similar. Building it costs a full extraction pass over the corpus, and what it buys is the ability to answer questions whose answer is a path or a pattern rather than a passage.