AI terms, pronounced.

A high-performance open-source runtime for running large language models on CPUs and GPUs, especially quantized GGUF models.

Wide Area Intelligence nodes can use llama.cpp-style local serving behind the gateway.

A local model runner that packages model download, storage, and serving behind a simple command-line and HTTP API.

People often use Ollama to try local models quickly before moving a model into a production node setup.

A desktop app for downloading, chatting with, and serving local models, commonly used on developer workstations.

LM Studio is a friendly way to test whether a machine can run a model before putting that hardware into a fleet.

A model file format used by llama.cpp and related runtimes. GGUF files usually include quantized model weights plus metadata.

If you are choosing a file from Hugging Face for a local node, GGUF is often the format you are looking for.

A compression technique that stores model weights with fewer bits, reducing VRAM and disk requirements at some quality cost.

A Q4 model is smaller and easier to run than an FP16 model, but may be less precise.

A common GGUF quantization preset. It is a practical default for many local LLMs because it keeps memory use low while preserving good output quality.

For many 7B to 14B models, Q4_K_M is the first file to try on consumer GPUs.

Video memory on a GPU. For local AI, VRAM is usually the hard limit that decides which model size and context window you can run.

Model weights, KV cache, and image/video generation pipelines all compete for VRAM.

Memory used by a transformer model to remember attention state for the current context. It grows as context length and parallel sessions grow.

A model can fit at 4K context but fail at 128K context because the KV cache becomes too large.

The maximum amount of text, images, tool output, and conversation history a model can consider in one request.

Coding agents often need larger context windows because files, instructions, and tool definitions consume tokens quickly.

A chunk of text used by a model. Tokens are not exactly words; short words may be one token, while long words can be several.

Cloud providers usually price text models per million input and output tokens.

A routing layer that gives applications one endpoint while deciding which model, provider, or local node handles each request.

Wide Area Intelligence is an LLM gateway that tries your hardware first and falls back to cloud when needed.

An API that follows the same request and response shapes as OpenAI's endpoints, so existing SDKs can point at another backend.

This lets apps switch their base URL to Wide Area Intelligence without rewriting their chat-completion code.

A fallback path that sends a request to a cloud provider when local hardware is offline, overloaded, or missing a required capability.

Failover keeps apps working while still letting local hardware serve the baseline workload.

A cache close to users that can return repeat responses without running inference again.

Exact-match AI cache hits are useful for repeated prompts, test suites, and deterministic internal workflows.

A secure outbound connection from private infrastructure to Cloudflare, used to expose services without opening inbound firewall ports.

It is one way to reach a private GPU without port forwarding. Wide Area Intelligence nodes instead open a plain outbound HTTPS connection — no inbound ports, and no separate tunnel daemon to run.

Low-Rank Adaptation: a smaller adapter trained to change a base model's behavior without retraining the full model.

LoRAs are common in image generation and increasingly used for lightweight specialization of language models.

A model architecture that activates only some expert subnetworks for each token, increasing total model capacity without using every parameter every time.

MoE models can be fast for their total parameter count, but still have memory and routing tradeoffs.

A pattern where an app retrieves relevant documents first, then includes them in the model prompt so answers can use external knowledge.

RAG is often better than fine-tuning when you need answers grounded in changing private data.

A numeric representation of text, images, or other data that makes similarity search possible.

Embeddings power semantic search, RAG retrieval, deduplication, and clustering.

Cloudflare's serverless AI inference platform, available from Workers code through an AI binding.

Wide Area Intelligence can use Cloudflare platform bindings alongside local node routing.

An edge-first LLM gateway that routes each request to hardware you own first, and fails over to the cloud only when needed.

Point an OpenAI-compatible app at Wide Area Intelligence and your own GPU node answers, with cloud failover as a backstop.

A large language model: a model trained on text to generate and reason over language, code, and structured output.

Local LLMs let you keep inference — and your data — on hardware you control.

Graphics Processing Unit: the parallel processor that runs most local AI inference. Its VRAM usually decides what you can run.

Wide Area Intelligence turns the GPU you already own into a node the gateway can route to.

Meta's family of open-weight language models, widely used as a base for local deployments and fine-tunes.

"Llama" is the model; "llama.cpp" is a runtime that can run it (and many other models).

A family of open-weight models from Alibaba, including strong coding-focused variants (Qwen-Coder).

Qwen-Coder models are a common local pick for coding agents behind the gateway.

A family of open-weight models known for strong reasoning and coding performance, often run locally as quantized GGUF files.

DeepSeek distillations are popular when you want reasoning quality on consumer hardware.

Google's family of small open-weight models, sized to run comfortably on modest GPUs.

Gemma is a common choice when VRAM is tight but you still want a capable general model.

Microsoft's family of small models designed to punch above their weight on reasoning and coding for the parameter count.

Phi models suit low-VRAM machines and short, well-scoped tasks.

A text command interface for a tool, driven by typed commands rather than a graphical window.

Ollama and many coding agents are driven from the CLI.

A graphical, windowed interface for a tool, as opposed to a command line.

LM Studio is a GUI app; llama.cpp is CLI-first.

An open-source terminal coding assistant that edits your local git repo through an OpenAI-compatible model endpoint.

Point Aider at a Wide Area Intelligence base URL and it codes against a model on your own GPU.

An autonomous coding agent (a VS Code extension) that plans and edits across files via an OpenAI-compatible model.

Cline is agent-heavy, so it wants a capable model and a large context window.

A family of open-weight language models from Zhipu AI, used as another local option for chat and coding.

GLM is one of the open-weight model families you can serve from a node behind the gateway.