Glossary

Accelerate is an open-source library from Hugging Face that makes PyTorch training code run on any hardware and in distributed setups with minimal changes — including mixed precision and FSDP and DeepSpeed support.

An adapter is a small, freshly added network module injected into a frozen language model — only these modules are trained, while the base model itself stays unchanged.

Agentic RAG is a RAG architecture in which an agent drives retrieval — it decomposes the question, searches iteratively, checks the hits, and decides itself whether a second or third pass is needed.

An AI agent combines a language model with tools and works toward a goal across multiple steps — typically in a loop of observing, planning and acting.

An AI content workflow is a structured process for creating content in which AI assists across several steps: briefing, draft, editing, fact-checking and publishing. The human steers, reviews and remains responsible for the result.

AI data analysis means evaluating data (CSV, Excel and more) via natural language: an AI system writes and runs code behind the scenes to clean, calculate and visualise files — e.g. ChatGPT's Advanced Data Analysis.

AI image generation creates images from text prompts using generative models. Tools such as Midjourney, DALL-E, Adobe Firefly and Stable Diffusion are used in marketing for concepts, visuals and image editing.

AI knowledge management makes scattered company knowledge searchable and usable with AI. Documents are turned into vectors so employees get answers from natural-language questions — usually via retrieval-augmented generation.

AI marketing is the use of artificial intelligence and machine learning to analyse marketing data, automate content and delivery, and optimise campaigns in real time. Its main areas are content, personalisation, analytics and generative engine optimization.

AI meeting assistants are software tools that automatically record, transcribe, summarize online meetings and derive action items from them. Well-known examples include Otter, Fireflies, Fathom and Microsoft Copilot in Teams.

AI presentations are slide decks created with the help of generative AI. From a short prompt or an outline, tools like Gamma or Beautiful.ai automatically generate structured slides with text, layout and matching images.

An AI research assistant is an AI system that, given a question, searches the web in multiple steps, reads and cross-checks sources, and synthesises a cited answer or report — e.g. ChatGPT Deep Research, Gemini Deep Research or Perplexity.

AI video generation creates moving images from text or image prompts. Diffusion-transformer models such as Google Veo, Kling, Runway or Sora iteratively denoise latent frames into coherent video, usually as clips of a few seconds.

AI-powered personalization automatically tailors content, offers and recommendations to individual users. Models analyse behaviour and profile data to deliver more relevant experiences in real time — consent-based and privacy-compliant.

Aider is an open-source AI pair programmer for the command line that edits code directly in the Git repository and commits every change automatically.

Universal-2 is AssemblyAIs closed speech-to-text model with industry-leading accuracy (mean WER around 6 percent) and strong formatting, punctuation and proper-noun recognition across 99 languages.

Computational method in transformer models that weights which parts of the input text are most relevant for predicting the next token — instead of treating every part as equally important.

Audio normalization is the adaptive linear scaling of an audio signal to a target level — the tool first measures a reference value (peak, RMS or LUFS) and from that calculates the required gain.

A safety layer in Claude Code: a small LLM evaluates every tool call that is not explicitly allowlisted, live, and blocks destructive or unauthorised actions before they execute.

AutoGen is an open-source framework from Microsoft for building multi-agent applications with LLMs. Several AI agents solve tasks together in structured conversations — for example group chats or asynchronous message passing.

AutoGPT is an open-source experiment released in 2023 that runs a language model such as GPT-4 in an autonomous loop — it breaks a goal into subtasks itself, calls tools and works through them with minimal human intervention.

Automatic Speech Recognition (ASR), also called Speech-to-Text (STT), is the automatic conversion of spoken language into written text by an acoustic model.

Asynchronous API mode that collects many requests and processes them at a significant discount — results are typically delivered within 24 hours.

Batch inference is the bundled processing of many prompts in one pass — locally on a GPU or as an asynchronous API job — trading real-time latency for throughput and cost.

An AI benchmark is a standardized task suite that makes language models comparable — same prompts, same scoring, transparent results for reasoning, knowledge, or coding.

Browser use refers to AI agents that drive a real web browser — navigating, clicking, filling forms, reading content — to complete web tasks autonomously. The browser serves as both the tool and the perception surface.

Catastrophic forgetting describes the effect that a neural network loses previously acquired knowledge — wholly or partially — when it is trained further on new data; a key problem when fine-tuning language models.

Chain-of-Thought (CoT) is a prompting technique that asks the model to spell out its reasoning in intermediate steps — boosting accuracy on multi-step tasks.

Chatterbox is an open-source text-to-speech model from Resemble AI under the MIT license. In blind tests most listeners preferred it over ElevenLabs; it offers emotion control, voice cloning and low latency.

Chroma (ChromaDB) is an open-source vector database for storing and searching embeddings. It is widely used in RAG systems to retrieve relevant text passages via similarity search and feed them to language models.

Chunking is the splitting of longer texts into smaller, self-contained pieces — the first step of any RAG pipeline because embeddings and retrieval work at the chunk level.

Claude Code is Anthropic's official AI coding tool — available as a terminal CLI, VS Code extension, JetBrains plugin and web app, with a strong focus on tool use and configurable workflows.

Claude Fable 5 is Anthropics first publicly available Mythos-class model, released in June 2026. It is built for the most demanding reasoning and long-horizon agentic work and offers a one-million-token context window.

Anthropic's flagship Claude Fable model, released on 1 September 2026. A frontier model for demanding reasoning, coding, research and long-running agents; 1M-token context, multimodal input, proprietary access via API and the Claude apps.

Claude Haiku 4.5 is Anthropics fast compact model from October 2025 — it delivers near-Sonnet-4 performance at roughly one third of the cost and is the first Haiku model with reasoning.

Claude Mythos (April 2026) is Anthropic's most powerful model to date — a new tier above the Opus line, distributed primarily through the Glasswing program to cybersecurity defenders.

Claude Mythos 5 (June 9, 2026) is Anthropic's most powerful model line above Opus — technically identical to Fable 5 but without its safety classifiers, distributed through Project Glasswing.

Claude Mythos 5.1 (September 2026) is Anthropics frontier model for vetted cyber and life-science organizations — the same model as Fable 5.1, only with more permissive safeguards.

Claude Opus 4.5 is Anthropic's top-tier model from November 2025 — the direct predecessor of Opus 4.6, bringing "Infinite Chats" and substantially reduced token usage.

Claude Opus 4.8 is the flagship model of Anthropics Opus line, released in May 2026. It targets agentic coding, complex reasoning and computer use, and is described as an incremental but measurable improvement over Opus 4.7.

Anthropic's flagship Claude Opus model, released on 24 July 2026. Positioned as a powerful everyday model for knowledge work, coding and agents; proprietary access via the Claude apps and API (model ID claude-opus-5).

Claude Sonnet 4.5 (September 2025) is Anthropics Sonnet-line model tuned for agents and coding — leading on SWE-bench Verified at launch, since superseded by newer Sonnet versions.

Claude Sonnet 4.6 is Anthropics mid-tier model from February 2026 — close to Opus-level performance, with a 1M-token context and unchanged Sonnet pricing of 3/15 USD per million tokens.

Claude Sonnet 5 is Anthropic's mid-tier model from June 30, 2026 — the most agentic Sonnet yet, with performance close to Opus 4.8 at $2 input and $10 output per 1M tokens.

Claude Sonnet 5.5 (September 28, 2026) is Anthropic's low-cost mid-tier model — a faster, cheaper companion to Opus 5.5, priced at $2/$10 per 1M tokens and leading on Terminal-Bench 4.0. Available via boostN.

Code Llama is Metas coding-specialized Llama 2 variant from August 2023 — open-weights, in several sizes and with base, Python and Instruct flavors.

Codestral is Mistral AI's code-specialized model (first released May 2024) — a 22B-parameter model for code completion and generation, originally published as open weights.

Computer Use is a capability of modern AI models that lets them operate a computer like a human — see the screen, move the mouse, use the keyboard — to perform tasks across arbitrary applications without an API.

Constitutional AI is a training method developed by Anthropic in which an AI model critiques and revises its own behaviour against an explicit list of principles (a "constitution") — rather than relying solely on human ratings.

Context engineering is the discipline of deliberately curating and maintaining everything inside the LLM context window — system prompt, tool definitions, RAG hits, memory, conversation history.

Context Precision is a RAG evaluation metric (0–1) measuring how well the retriever ranks relevant contexts above irrelevant ones — the focus is ranking quality, not raw hit count.

Context Recall is a RAG evaluation metric (0–1) measuring how completely the retriever supplied the facts needed for the gold answer — every missing fact lowers the score.

The context window is the maximum number of tokens a language model can process at once — input and output combined.

Contextual Retrieval is an Anthropic RAG technique that prepends an LLM-generated context sentence to every chunk before indexing — sharply raising retrieval quality.

Continue.dev is an open-source, model-agnostic AI coding extension for VS Code and JetBrains — configured via YAML, working with commercial as well as local models.

Continued pretraining is the further training of an already pretrained language model on large amounts of domain text — before any classic fine-tuning begins.

Conversational marketing is a dialogue-based form of customer engagement that responds to prospects in real time via chat, messaging and AI chatbots. The goal is to guide users along the customer journey through conversation rather than rigid forms.

Cosine similarity measures how similar two vectors are by the angle between them — the default metric in RAG retrieval for comparing query embeddings against the embeddings stored in a vector index.

CrewAI is an open-source framework for orchestrating multiple autonomous AI agents. It assigns agents roles, goals and tools and has them collaborate in teams ("crews") to solve complex tasks through coordinated, multi-step workflows.

Cursor is an AI-first code editor — a fork of Visual Studio Code with AI deeply wired into the editor itself.

Datasets is an open-source Python library from Hugging Face for loading, processing and sharing machine-learning datasets. It uses Apache Arrow and memory-mapping to handle even very large amounts of data efficiently.

Three related concepts for digital audio levels: dBFS is the scale with 0 dBFS as the hard ceiling, headroom is the safety margin below the maximum, clipping is the hard cut-off when a signal exceeds it.

Nova-3 is Deepgrams real-time speech-to-text model (released February 2025) with very low streaming latency of around 200 to 300 milliseconds, keyterm prompting and support for more than 30 languages.

Janus-Pro-7B is DeepSeek's open-source multimodal model from January 2025 — unifying image understanding and image generation in a single 7B architecture, MIT-licensed.

Open reasoning model from the Chinese lab DeepSeek, released on 20 January 2025 under the MIT license. A mixture-of-experts with 671B parameters (37B active per token), heavily trained via reinforcement learning; the first open model to match OpenAI's o1 level.

DeepSeek V3 is an open model from the Chinese provider DeepSeek, released in December 2024. It uses a Mixture-of-Experts architecture with 671 billion parameters (37 billion active per token) and a context window of around 128,000 tokens.

DeepSeek V3.1 is DeepSeeks open-weights hybrid model from August 2025 — it unifies a think and a non-think mode in a single model, with a 128K context and strong agentic skills.

DeepSeek V3.2 (December 1, 2025) is an open MoE model with 685B parameters under an MIT license — especially efficient on long contexts thanks to DeepSeek Sparse Attention.

DeepSeek V4 (April 2026) is the fourth generation of the Chinese open-weight MoE model — released as Pro (1.6T parameters) and Flash (284B) under MIT license, with a 1M-token context window.

DeepSeek V4-Flash (July 2026) is an open MoE model with 284B parameters (13B active), a 1M-token context and an MIT license — freely downloadable on Hugging Face.

DeepSeek V4-Pro (2026) is DeepSeeks open-weight flagship — a 1.6-trillion-parameter MoE with about 49B active parameters per token, a 1M-token context and an MIT license.

DeepSeek V4.1-Flash (September 10, 2026) is the smallest model in the new DeepSeek V4.1 architecture family — the first DeepSeek generation with native multimodal image processing. A speed-/cost-optimized "Flash" model with open weights (MIT), available via boostN.

Devstral is Mistral AI's open agentic model for software engineering (May 2025) — trained to solve real GitHub issues and strong on SWE-bench Verified.

Diffusers is an open-source Python library by Hugging Face for using and training diffusion models for image, video and audio generation. It is built on PyTorch and is structured into pipelines, models and schedulers.

DPO is a fine-tuning method that aligns language models directly to human preference pairs — without a separate reward model and without reinforcement learning.

DSPy is an open-source Python framework from the Stanford NLP group that lets you program LLM applications instead of hand-crafting prompts. Tasks are described through declarative signatures and modules; optimizers automatically derive effective prompts and examples.

ElevenLabs v3 is a highly expressive text-to-speech model with inline audio tags for emotion, emphasis and non-verbal sounds. It supports more than 70 languages and multi-speaker dialogue and is seen as a benchmark for realism.

An embedding is a numeric representation (vector) of text, an image, or other data in which semantically similar content sits close together in space.

Ensemble means combining several deliberately varied LLM runs or models whose findings complement each other. Multi-model orchestration drives these runs via orchestrators with sub-agents, so the union of results is larger than any single run.

Evals are systematic tests for LLM applications — fixed test cases, automatic scoring of the responses, result as a score. The foundation for objectively comparing prompt or model changes.

Faithfulness is a RAG evaluation metric (0–1) measuring how many claims in a response can actually be derived from the retrieved contexts — a direct hallucination indicator.

Few-shot prompting is a technique that includes a handful of input/output examples in the prompt so the model picks up the desired format and style.

Continued training of a pretrained model on a smaller, specialized dataset to adapt style, format or domain knowledge in a targeted way.

Fish Audio S2 is an open-source text-to-speech model from the Fish Speech family. With inline tags it controls emotion and emphasis at word level, supports more than 80 languages and delivers very low latency.

Flowise is an open-source low-code tool for building LLM applications and AI agents visually via drag-and-drop. It is built on the LangChain ecosystem and connects models, data sources and tools as nodes.

FLUX.2 is the text-to-image model family from Black Forest Labs. It spans the proprietary Pro and Flex variants plus the open 32-billion-parameter Dev model and the compact Klein series, unifying image generation and image editing in a single model.

Foundation-Sec-8B is Cisco Foundation AIs open, cybersecurity-specialized language model (2025). It extends Llama-3.1-8B via continued pretraining on about 5 billion security tokens and is available under the permissive Apache-2.0 license.

Function calling is a language model's ability to produce a structured function invocation instead of a text reply — the technical foundation for tool use and AI agents.

Gemini 3 Flash is Google's fast, low-cost thinking model from December 17, 2025 — near-Pro reasoning with a 1M-token context, built for agentic workflows and coding.

Gemini 3 Pro (November 2025) was Google's third flagship model with a 1M-token context window and native multimodality — now superseded by Gemini 3.1 Pro.

Gemini 3.1 Pro (February 2026) is Google's current flagship — 1M-token context, 64K output tokens, doubled reasoning performance over Gemini 3 Pro.

Gemini 3.5 Flash is Googles high-efficiency multimodal model from May 2026 — it brings near-Pro coding and reasoning quality at Flash-typical cost and speed.

Gemini 3.6 Flash (July 2026) is Googles fast, cost-efficient Flash tier with a 1M-token context — built for high throughput and low cost per request.

Gemini 3.7 Flash (August 2026) is Googles fast workhorse model for coding and agents — 1M-token context, introductory price of 0.75/3.75 USD per million tokens, half the cost of 3.6 Flash.

Gemini 3.8 Flash (September 2026) is Googles most intelligent workhorse model for coding, agents and multi-step reasoning — 1M-token context, introductory price of 0.75/3.75 USD per 1M tokens, direct successor to 3.7 Flash.

Gemma 3 is Googles family of open models released in March 2025. It comes in four sizes (1B, 4B, 12B, 27B), is multimodal from 4B upward, supports over 140 languages and offers a context window of up to 128,000 tokens.

Google's open, mobile-first AI model, available in preview from May 2025 and fully from 26 June 2025. Multimodal (text, image, audio, video), optimised for on-device use; open weights in the E2B and E4B sizes.

Gemma 4 is Google DeepMind's open model family from April 2026 — four sizes (E2B to 31B), 256k context, multimodal, Apache 2.0 licensed.

GGUF (GPT-Generated Unified Format) is the standard file format for quantised LLMs on the llama.cpp engine — a single `.gguf` file holds weights, tokenizer and metadata together.

GitHub Copilot is GitHub's AI coding assistant — an extension for common editors offering tab autocomplete, chat and multi-file edits, with GPT and Claude models selectable.

GLM-4.5 is an open model from Zhipu AI (Z.ai) released in July 2025. It is an agent-native Mixture-of-Experts model with 355 billion parameters (32 billion active), a 128,000-token context and an MIT licence.

GLM-4.6 is Zhipu AI's open-weight MoE model from September 2025 — 355B parameters (32B active), 200k context, MIT-licensed, with a focus on coding performance.

GLM-4.7 (December 2025) is Zhipu AIs open-weight coding flagship — an MoE model with roughly 400B parameters, a 200K input context and the "Preserved Thinking" feature.

Open model from Zhipu AI (Z.AI), released in February 2026 under the MIT license. A mixture-of-experts with around 744B parameters (about 40B active per token), 200K context, focused on coding and agents. Freely downloadable via Hugging Face.

GLM-5.1 is Z.ai's open-source flagship from April 2026 — 744B parameters (40B active), 200k context, designed for agentic coding with up to 8 hours of autonomous runtime.

GLM-5.2 (June 2026) is Zhipu AIs (Z.ai) open-weight model — an MoE with around 744B parameters (~40B active), a 1M-token context and MIT license, strong at agentic coding.

GLM-5.3 is Zhipu AIs open-weights model from August 2026 — a MoE model focused on coding and cybersecurity that tops the open-model field on agentic benchmarks.

Google Imagen is the text-to-image model family from Google DeepMind. The current generation, Imagen 4, launched in 2025 in Fast, Generate and Ultra tiers, is available via the Gemini API and Vertex AI, and is known for strong typography and prompt adherence.

GPT-6 Astra is OpenAIs flagship from September 2026 — the first model at the Critical Preparedness tier, focused on computer use and cybersecurity.

GPT Image 2 is OpenAIs native image model (April 2026) that reasons before drawing, renders text very reliably and produces high-resolution photorealistic images.

GPT Luna (GPT-5.6 Luna) is the smallest, fastest and most affordable model of OpenAIs GPT-5.6 family, released in July 2026. It is built for cost-sensitive, high-volume workloads and offers a context window of roughly 1.05 million tokens.

GPT-5.6 Sol is OpenAIs flagship from July 2026 — the top model of a new solar-system family (Sol, Terra, Luna) focused on coding, science and cybersecurity.

GPT Terra is the middle tier of OpenAIs GPT-5.6 family (July 2026), sitting between the efficient Luna and the flagship Sol — built for a strong price-performance balance with a 1M-token context.

GPT-4o (May 2024) was OpenAI's first natively multimodal model — text, audio and image in a single model, with a real-time voice mode and substantially lower cost than GPT-4 Turbo.

GPT-4o mini (July 2024) is OpenAI's small, low-cost variant of GPT-4o — positioned as the successor to GPT-3.5 Turbo, clearly stronger at a much lower price per token.

GPT-5.2 is OpenAI's model generation from December 2025 — three variants (instant, thinking, Pro), a 400k context window and new state-of-the-art on professional knowledge work.

GPT-5.3 is an iteration of OpenAI's GPT-5 line — the Instant variant launched on March 3, 2026 with a focus on factual accuracy, shorter preambles and a more natural conversational flow.

GPT-5.3-Codex is OpenAI's specialized agentic coding model (February 2026), the successor to GPT-5.2-Codex — 25% faster, with substantial gains on Terminal-Bench and OSWorld.

GPT-5.3-Codex-Spark is OpenAI's coding model from February 2026 — a smaller, real-time-optimized variant of GPT-5.3-Codex delivering over 1000 tokens per second on Cerebras hardware.

GPT-5.4 is OpenAI's reasoning model from March 5, 2026 — the first mainline model with the coding capabilities of GPT-5.3-codex, native computer-use controls and 1M-token context.

GPT-5.5 is OpenAI's flagship from April 23, 2026 — positioned as its "smartest and most intuitive model" and pitched as a step toward AI that works across tools on its own.

GPT-6.1 Sol (September 29, 2026) is OpenAIs low-cost coding and agentic model — the successor to GPT-6 Sol after just seven days, with near GPT-6 Astra intelligence at roughly a fifth of the price. Available via boostN.

OpenAI's real-time voice model, the default voice experience in ChatGPT since 8 July 2026. It listens and speaks at the same time (bidirectional), allows interruptions, and hands complex questions off to a stronger text model. Proprietary access via ChatGPT.

GPT4All is an open-source desktop application by Nomic AI for running large language models locally and offline on your own computer. It works without a GPU and without a cloud connection, so data never leaves the device.

GraphRAG is a RAG variant that uses a knowledge graph in addition to vector search — answers come from linked entities and relationship paths, not just similar text.

Reasoning model from xAI (Elon Musk's AI company), released on 17 February 2025. Trained on the Colossus supercluster, with 'Think' reasoning and DeepSearch plus real-time access to X. Proprietary access via the Grok app and API.

Grok 4 is xAI's reasoning model from July 2025 — text and image input, 256k context, trained via reinforcement learning on the 200,000-GPU Colossus cluster.

Grok 4.1 is xAIs AI model released in November 2025. It delivers more natural dialogue, fewer hallucinations and better style control, and comes in two variants: one that answers directly and one that reasons before responding (Thinking).

Grok 4.3 is xAIs reasoning-oriented model from April 2026 — with a 1M-token context, native video input and a focus on agentic workflows and factual accuracy.

Grok 4.5 is xAI's July 2026 frontier model for coding, agentic tasks and knowledge work — high token efficiency at $2 input and $6 output per 1M tokens.

Grok 4.6 (August 2026) is xAIs flagship model for long-running agentic work — 500k-token context, text and image input, tiered pricing from 2/6 USD per million tokens.

Guardrails are protective layers around a language model that inspect inputs and outputs to catch unwanted behaviour — off-topic answers, PII leaks, unsafe actions.

A hallucination is a language-model response that sounds plausible but is factually wrong or fabricated — typically caused by statistical guessing without a factual anchor.

Haystack is an open-source AI orchestration framework from deepset for building production-ready LLM applications. Modular components — retrievers, generators, routers, tools — are assembled into explicit pipelines for RAG, agents and semantic search.

Hugging Face is the central platform of the open-source AI community — a hub for models, datasets, and demos, and the maintainer of widely used libraries like Transformers, Diffusers, and PEFT.

Central platform for openly shared AI models, datasets and demos — a kind of GitHub for machine-learning artifacts.

Hunyuan3D 2.0 is Tencents open generative 3D model (January 2025). In two stages it produces high-resolution, textured 3D assets from text or image and supports text-, image-, sketch- and portrait-to-3D. The weights are freely available on Hugging Face.

Hybrid search combines lexical full-text search (e.g. BM25) with semantic vector search and unites the strengths of both — precise on keywords and robust on meaning.

HyDE (Hypothetical Document Embeddings) is a RAG technique where an LLM first drafts a fake answer to a query — and then uses that answer's embedding to find real documents.

Hyper3D Rodin is a commercial 3D generation model by Deemos. Generation 2.5 turns images or text into highly detailed geometry with over 10 million polygons and 3D-native PBR textures, and is regarded as a leader in photorealism.

Indirect prompt injection is an attack that hides malicious instructions inside external content — web pages, documents, emails — that an LLM then processes and unintentionally executes.

Inference is the act of running an already-trained model to turn an input into a response — production use, not training.

Tokens you send to an AI model in an API call — your prompt, the context, attached documents. Billed separately from output tokens and usually much cheaper.

Instruction tuning is a variant of supervised fine-tuning in which a language model is trained on a broad mix of instruction-response pairs — so it reliably follows instructions across arbitrary tasks.

A jailbreak deliberately bypasses a language model's safety and behaviour rules through cleverly crafted prompts, making the model produce content it would normally refuse.

Jan is an open-source desktop application that lets you run large language models locally and fully offline on your own computer. It uses the llama.cpp engine and is positioned as a privacy-friendly ChatGPT alternative.

A keyword glossary for Speech-to-Text is a short, curated list of project-specific terms, brand and product names handed to an STT model as a context hint before transcription — typically through Whisper's `initial_prompt`.

Kimi K2 is Moonshot AI's open-weight MoE model series with 1T total parameters and 32B active parameters, trained on 15.5T tokens — focused on agentic and coding tasks.

Kimi K2 Thinking (November 2025) is Moonshot AI's open reasoning-agent model — a trillion-parameter MoE with a 256K context that runs long thinking and tool-use chains on its own.

Kimi K2.5 is an open model from Moonshot AI released in January 2026. It is a native multimodal Mixture-of-Experts model with around 1 trillion parameters (32 billion active), a 256,000-token context and agentic capabilities.

Kimi K2.6 is Moonshot AI's open-weight agent model from April 2026 — 1T parameters (32B active), 262k context, natively multimodal, with agent swarms of up to 300 sub-agents.

Kimi K3 (July 2026) is Moonshot AIs open-weight flagship — an MoE model with 2.8T parameters (104B active), native vision and a 1M-token context.

Kling 3.0 is Kuaishous video model, released on February 5, 2026. It generates photorealistic clips of up to 15 seconds with native audio across multiple languages and is regarded as the strongest price-to-performance video generator.

Knowledge distillation is a training technique in which a small student model learns from the behaviour of a large teacher model — aiming for comparable quality at significantly lower resource cost.

Kokoro is an open text-to-speech model (v1.0, January 2025) with just 82M parameters. It produces natural-sounding speech in 8 languages with 54 voices, runs on around 1 GB of VRAM and ships under the Apache-2.0 license.

The KV cache (key-value cache) stores the already computed key and value vectors of a language model's attention layers during text generation. This avoids recomputing the entire context for each new token and significantly speeds up inference.

Open-source framework that chains LLMs with data sources, tools and memory into applications — arguably the best-known pipeline library for AI apps.

LangGraph is an open-source framework from LangChain for stateful, graph-based LLM workflows — nodes are functions or agents, edges define the control flow including loops and branches.

Llama 2 is Metas open-weights language model from July 2023 — the first Llama with a commercially usable license, in 7B, 13B and 70B sizes with a 4K context.

Meta's open model generation, released on 18 April 2024. It launched in the 8B and 70B sizes (each in base and Instruct variants) under the Llama 3 community license; freely downloadable, for research and commercial use.

Llama 3.1 (July 2024) is Metas open-weight family in 8B, 70B and 405B with a 128K context and multilinguality — the 405B was the first open model on par with GPT-4o and Claude 3.5 Sonnet.

Llama 3.2 is Meta's model generation from September 2024 — the first with vision support (11B/90B) and lightweight edge models (1B/3B) for mobile and on-device use.

Llama 3.3 (December 2024) is Metas open 70B instruct model — it reaches 405B-class quality at 70B cost, with a 128K context and open weights.

Llama 4 Maverick (April 2025) is Meta's large Mixture-of-Experts variant in the Llama 4 lineup — 17B active out of 400B parameters across 128 experts, 1M-token context.

Llama 4 Scout (April 2025) is Meta's smaller Mixture-of-Experts variant in the Llama 4 lineup — 17B active out of 109B parameters, 10M-token context, multimodal.

Llama-3-SauerkrautLM-70b is a German-language DPO fine-tune of Meta Llama 3 70B, developed by VAGOsolutions and Hyperspace.ai.

llama.cpp is an open-source C++ library for efficient LLM inference on CPU and GPU — the engine many local tools like Ollama and LM Studio are built on.

Llamafile is a Mozilla project that packages a local language model as a single executable file — model weights and inference engine in one. That file runs on several operating systems without installation.

Data framework for LLM applications, specialized in indexing, enriching and retrieving private data — at the core of many RAG architectures.

LLM stands for Large Language Model — a neural network trained on large volumes of text to interpret and generate natural language.

LLM-as-a-Judge is the practice of using a language model as an evaluator — it compares responses from other models or scores outputs against given criteria.

LM Studio is a desktop application for running LLMs locally with a graphical interface — including a model browser, chat interface and OpenAI-compatible server.

Low-Rank Adaptation — parameter-efficient fine-tuning that trains only small additional matrices instead of all model weights.

Magistral is Mistral AIs first reasoning model from June 2025 — with a transparent, verifiable chain of thought and strength in European languages, partly as open weights.

Make (formerly Integromat) is a cloud-based no-code workflow automation platform. Through a visual interface, you connect apps and APIs into multi-step workflows ("scenarios") without programming — for example to pass data between services.

MCP — Model Context Protocol — is an open standard from Anthropic for connecting AI models to external data sources and tools through a unified interface.

Meshy is a commercial hosted platform for 3D generation from text or images. Version 6 produces production-ready assets in about a minute with PBR textures, auto-retopology and rigging, and is regarded as a versatile all-round tool.

Meta-prompting is the technique of having an LLM write or improve a prompt for a task rather than formulating it yourself — the prompt becomes the output.

Microsoft TRELLIS.2 is the strongest open image-to-3D model (4B parameters, MIT license). It generates high-resolution assets with native PBR materials in seconds and supports Gaussian Splatting; weights are freely available on Hugging Face.

Midjourney v8 is the eighth generation of Midjourneys image generator (alpha March 2026) with roughly five times faster generation, native 2K resolution and improved text rendering.

Milvus is an open-source, distributed vector database for fast similarity search over large sets of embeddings. It is licensed under Apache 2.0 within the LF AI & Data Foundation; the main developer is the company Zilliz.

MiniMax is an AI lab founded in Shanghai in 2021 whose open-weight M model family (LLMs) targets very long context windows and low running costs via sparse attention.

Mistral 7B (September 2023) is Mistral AIs first open-weight model — a 7.3B-parameter model under Apache 2.0 that at the time outperformed the larger Llama 2 13B on benchmarks.

Mistral Large 3 (December 2025) is Mistral's open-weight MoE flagship — 675B parameters, 256K-token context, natively multimodal, Apache 2.0.

Mistral Medium 3 is a proprietary AI model from Mistral AI released in May 2025. As an enterprise model built for high performance at low operating cost, it offers around 128,000 tokens of context and can also be run on-premise or in your own VPC.

Mistral Small 4 is Mistral AI's open-weight MoE model from March 2026 — merging reasoning, vision and coding into one model, 119B parameters (6.5B active), Apache 2.0.

Mixtral 8x7B (December 2023) is Mistral AIs open Sparse-MoE model with 8 experts — 46.7B parameters, 12.9B active per token, 32K context, Apache 2.0 license.

A model card is a standardised datasheet for an AI model — documenting purpose, training data, performance, limitations and ethical considerations in one place.

Muse Spark is Metas closed-weight flagship from April 2026 — the first model of the Muse line out of Meta Superintelligence Labs, focused on efficient reasoning via thought compression.

Canary is NVIDIAs open ASR and speech-translation model family (CC-BY-4.0) that tops the Open ASR Leaderboard on English accuracy ahead of OpenAI Whisper and covers 25 European languages.

Ollama is a tool for running language models locally — Llama, Mistral or Qwen — operated via a CLI, with an OpenAI-compatible HTTP server for applications.

Open WebUI is an open-source chat interface for local and remote LLMs — a ChatGPT-like browser UX that runs as a Docker container and talks to Ollama, OpenAI-compatible APIs and more.

Claude Opus 4.6 is Anthropic's flagship model from February 2026 — top tier of the Claude 4 family with a 1M-token context window, stronger coding and new effort controls.

Claude Opus 4.7 is Anthropic's top-tier model from April 2026 — successor to Opus 4.6 with high-resolution image processing, a new xhigh effort level and task budgets for long agentic runs.

ORPO is a training method that unifies supervised fine-tuning and preference optimisation in a single step — leaner than the classic SFT-plus-RLHF pipeline.

Tokens an AI model produces as its response. Billed separately and usually three to five times more expensive than input tokens because the model has to actively generate them.

Path dependency describes how, during LLM generation, each early token shapes everything that follows — the first verbalized finding steers the entire remaining analysis in one direction.

PEFT bundles methods that adapt a pretrained LLM by training only a small fraction of its weights — typically under 1%, instead of all billions of parameters.

A global setting in Claude Code that controls how tool calls are handled: Plan (planning only), Default (asks for confirmation), Accept Edits (files free), Bypass (allow everything).

pgvector is an open-source PostgreSQL extension for vector search — adding embedding columns, distance functions, and ANN indexes (HNSW, IVFFlat) directly to an existing Postgres database.

Pinecone is a managed vector database service for RAG and semantic-search workloads — it stores embeddings, answers nearest-neighbour queries, and scales without the user operating the index themselves.

Mistral AI's first multimodal model, released in September 2024 under Apache 2.0. A vision-language model with 12B parameters plus a 400M vision encoder, processing text and (multiple) images; 128K context, open weights via Hugging Face.

Prefix tuning is a PEFT method that trains a sequence of learned "virtual tokens" prepended to the input rather than any model weights — the base model itself stays frozen.

Prompt caching is an API feature in which a provider stores recurring prompt prefixes — making subsequent requests cheaper and faster because the cached portion is not reprocessed.

Prompt engineering is the discipline of steering AI models through carefully crafted input prompts — closer to precise writing than to classic programming.

Prompt injection is an attack that hides instructions inside an LLM's input data to make the model ignore or subvert its original instructions.

Prompt leaking is an attack that tricks an LLM into revealing its hidden system prompt or other confidential context contents — a special case of prompt injection.

A prompt library is a central, organized collection of reusable and versioned prompts within a team. It provides proven inputs for AI models so they do not have to be reworded every time they are used.

A prompt template is a reusable prompt skeleton with placeholders that get filled with concrete values at runtime — the basis for reproducible LLM calls in applications.

Qdrant is an open-source vector database written in Rust — it stores embeddings for semantic search, RAG and recommendation systems and runs locally, on your own cluster or as a managed cloud.

Extension of LoRA that quantizes the base model to 4 bits — allowing even very large LLMs to be fine-tuned on a single GPU.

Quantization reduces the numeric precision of model weights — e.g. from 16-bit float to 4-bit integer — making language models small and fast enough for consumer hardware, with manageable quality loss.

Query expansion enriches the original query in a RAG pipeline with synonyms, reformulations or hypothetical answer texts in order to lift retrieval recall.

Open model family from Alibaba, released in September 2024. It spans base sizes from 0.5B to 72B parameters plus Coder and Math variants; 128K context, multilingual. Smaller sizes under Apache 2.0, the 72B model under a dedicated Qwen license.

Qwen 3 is Alibabas family of open models released in April 2025. It ranges from 0.6B to 235B parameters (dense and MoE models), is licensed under Apache 2.0 and unites a thinking and a non-thinking mode in a single model.

Qwen 3.5 (February 2026) is Alibaba's third-generation open model family — from 0.8B to 397B parameters, MoE flagship with 1M-token context, multimodal.

Qwen 3.6 is Alibabas open-weights model from April 2026 — the dense 27B variant matches much larger MoE models on agentic coding benchmarks, under an Apache 2.0 license.

Qwen 3.7-Max (May 2026) is Alibabas proprietary Max flagship with a 1M-token context, positioned as the "Agent Frontier" for long, tool-intensive agent workflows.

Qwen 3.8-Max (August 2026) is Alibabas largest flagship — an MoE model with 2.4T parameters, a 1M-token context and multimodal input, later released with open weights.

Qwen3-Max is Alibaba's trillion-parameter flagship from September 2025 — a Mixture-of-Experts model with a 262K context that, unlike earlier Qwen models, is API-only.

RAG (Retrieval-Augmented Generation) connects a language model to an external knowledge source — relevant passages are retrieved and passed to the model alongside the question.

RAGAS is an open-source framework for automated evaluation of RAG and agent pipelines — with built-in metrics like faithfulness, context precision, and answer relevancy.

Provider-enforced cap on requests or tokens per time window — it protects infrastructure and ensures fair usage across customers.

ReAct is a prompting pattern that has an LLM alternate between reasoning (thoughts) and action (tool calls) — the foundation behind many agent implementations.

Reasoning Effort is a control parameter on modern reasoning models that sets how much internal step-by-step thinking (thinking tokens) a model spends before answering — higher levels raise quality but also latency and cost.

Recall (hit rate, sensitivity) measures the share of all actually existing relevant cases that a system finds. Formula: relevant cases found divided by all relevant cases that truly exist. Requires a known ground truth.

Reranking reorders an already-retrieved list of hits with a more accurate model — typically a cross-encoder that scores query and each candidate together, instead of just comparing vector distances.

Reinforcement Learning from Human Feedback — training procedure that aligns models with helpful and safe behavior using human preference comparisons.

Role prompting assigns the LLM a concrete role or persona ("you are an experienced tax lawyer …") to steer style, vocabulary and depth of answers.

Runway Gen-4.5 is Runways video model that turns text or an image into 5- and 10-second clips with strong prompt adherence and cinematic motion. It was built with NVIDIA and uses an Autoregressive-to-Diffusion technique.

Safetensors is a file format developed by Hugging Face for storing model weights — faster, safer and more language-agnostic than the older PyTorch `.pt`/`.bin` format.

Sampling is the weighted drawing of the next token from a language model's probability distribution — governed by temperature, top-p and top-k. It produces the variability between two runs.

SauerkrautLM is a German-language LLM family from the German startup VAGOsolutions — fine-tunes based on Llama, Qwen, Mistral, and other open architectures.

Sec-Gemini v1 is Googles experimental, Gemini-based AI model for cybersecurity. It combines Gemini reasoning with current threat intelligence from Mandiant, the OSV vulnerability database and Google Threat Intelligence.

Seedream 4.0 is ByteDances multimodal image model (September 2025) that processes text and multiple images as input, renders up to 4K and runs more than ten times faster than Seedream 3.0.

Self-Consistency is a prompting technique where the same question is answered multiple times with chain-of-thought, and the most frequent answer is picked as the final result.

Self-Refine is a prompting technique in which a model critiques its own output and improves it across several iterations — without human feedback.

Semantic search retrieves content by meaning rather than exact word match — query and documents are compared as embeddings, so synonyms and paraphrases also hit.

SFT is the supervised post-training of a pretrained language model on a dataset of input-output pairs — the step that turns a base model into a usable assistant.

Hugging Face Spaces is a hosting platform for deploying and sharing interactive machine learning applications without your own server infrastructure. Apps become usable directly in the browser, for example via Gradio, Streamlit, Docker or static HTML.

Speech-to-Text refers to the automatic conversion of spoken language into text by an AI model. In an AI workflow, STT replaces the keyboard as the input channel — what matters is model size and domain vocabulary.

Spotlighting is a defense technique against prompt injection in which untrusted inputs are marked so the model treats them as data — not as instructions.

Stop sequences are strings that make an LLM end token generation as soon as they appear. They constrain the output deliberately — to prevent role switches, format markers or overly long answers.

Streaming means transmitting an LLM's response token by token in real time — the user sees the text word by word rather than only after the full generation completes.

Structured output is a language model's ability to deliver responses in a defined schema (typically JSON), reliable enough to be consumed directly by downstream code.

Synthetic data is artificially generated training or evaluation data — typically produced by a language model itself to extend datasets without needing real-world sources.

A system prompt is the instruction that shapes a language model's behavior across an entire conversation — in contrast to the user prompt, which varies per message.

Temperature is a sampling parameter that controls how deterministic or creative a language model's responses are — low values stay conservative, high values produce more variety.

TensorRT-LLM is an open-source library from NVIDIA that optimizes LLM inference on NVIDIA GPUs. It compiles models into highly optimized runtime engines and offers in-flight batching, a paged KV cache and quantization for high throughput at low latency.

Text-Generation-WebUI (oobabooga) is an open-source web interface for running LLMs locally. It bundles several inference backends, chat and completion modes and an OpenAI-compatible API under one common UI.

TGI (Text Generation Inference) is an open-source toolkit from Hugging Face for serving large language models efficiently. It provides a production-ready inference server with continuous batching, token streaming and optimised attention methods.

The thinking budget (token budget) is the maximum number of internal reasoning tokens a reasoning model may spend on "thinking" per request — controlling the trade-off between answer depth and cost or latency.

Tiered model line-up from a provider — small fast variants (Mini/Flash/Haiku) at a fraction of the price of the big frontier models. Also: volume tiers with quantity discounts.

A token is the smallest unit a language model works with internally — usually a sub-word fragment, occasionally a single character.

A tokenizer is the program that splits text into tokens before a language model can process it — it determines how many tokens a piece of text costs.

An invocation of a tool by an AI model during a conversation — such as reading a file, running a bash command, fetching from the web or calling an MCP tool. The foundation of agentic workflows.

Top-p (nucleus sampling) and top-k are sampling strategies for LLMs that decide which token candidates to draw from at each step — together with temperature they control the creativity of the output.

TPM (tokens per minute) and RPM (requests per minute) are the two common units in which AI API providers express their rate limits — TPM caps the token volume, RPM caps the number of requests per minute.

Open-source Python library from Hugging Face that gives unified access to thousands of pretrained models across text, vision and audio.

Tree of Thoughts (ToT) is a prompting technique in which an LLM explores multiple solution paths as a branching tree, scores them and follows only the promising branches — a generalisation of Chain-of-Thought.

Tripo AI is a commercial 3D generation platform by VAST AI. Version 3.1 leads the 2026 ELO arena rankings and turns text or images into game-ready models with clean quad topology, PBR textures and auto-rigging.

A vector database stores embeddings as high-dimensional vectors and searches them by semantic similarity — the core infrastructure for RAG, semantic search, and recommendation systems.

Veo 3.1 is Googles video model (DeepMind) that turns text or an image into 8-second clips with natively synchronized audio. Since the January 2026 update it delivers true 4K (3840x2160) and native vertical formats.

vLLM is an open-source high-throughput inference server for LLMs — known for PagedAttention, continuous batching, and an OpenAI-compatible API endpoint, widely used in production self-hosting setups.

VRAM (Video RAM) is the dedicated memory on a GPU — the single most important hardware figure for running language models locally, because ideally the whole model and its context must fit inside it.

Weaviate is an open-source vector database written in Go — it stores objects together with their embeddings, combines semantic and keyword (BM25) search and offers built-in vectorization, RAG and multi-tenancy.

Whisper is OpenAI's open speech-to-text model from 2022 — a multilingual encoder-decoder transformer that turns audio into text and ships in several sizes under an MIT license.

The `initial_prompt` parameter lets you pass Whisper a keyword list or sample sentence before transcription. The model treats this text as context and recognises the included terms far more reliably.

whisper.cpp is an open-source C++ port of OpenAI's Whisper speech recognition — a compact native binary for local transcription on CPU, CUDA, Metal or Vulkan, with no Python dependency.

Windsurf is an AI IDE built on the VS Code base with a focus on agentic coding workflows — the Cascade agent plans and executes multi-step changes autonomously.

XML tags are named brackets like <context>...</context> used to clearly separate sections of a prompt. They structure the input, split instruction from data and make the model's output more predictable.

Zapier AI bundles the AI features of the automation platform Zapier — including Copilot for building workflows by language, autonomous agents, chatbots and an MCP server that exposes thousands of Zapier actions to external language models.

Zero-shot describes solving a task with a language model without providing any examples — purely via the task description in the prompt.