Glossary

Key terms and concepts used across the Archipelag.io platform.

Glossary

Key terms used across the Archipelag.io platform, documentation, and community.

Island

A compute contributor on the Archipelag.io network. An Island is any device — desktop, server, laptop, or phone — that runs the Archipelag.io Island software and makes its GPU or CPU available to the network. Think of each contributor as an island in the archipelago: independent, but connected.

Examples:

• “Your Island earned 42 credits today.”
• “3 Islands are online in your region.”
• “Set up your Island in under 5 minutes.”

Previously known as: Host, Node Agent

Consumer

Someone who uses compute on the Archipelag.io network. Consumers submit jobs — like chat completions or image generation — and the platform routes them to nearby Islands. Think of a consumer as someone sailing between islands, picking up what they need.

Examples:

• “Consumers pay per job at market rates.”
• “As a consumer, your requests are routed to the nearest available Island.”

Coordinator

The Archipelag.io control plane. It handles job routing, Island registration, billing, reputation scoring, and real-time streaming. Consumers and Islands never talk directly — everything flows through the Coordinator.

Built with: Elixir, Phoenix, PostgreSQL

Job

A single unit of work submitted by a consumer. A job might be a chat completion, an image generation request, or a batch inference task. Jobs are dispatched to Islands, executed, and results are streamed back.

Lifecycle: pending → assigned → running → completed (or failed)

Cargo

A Cargo is a pre-approved, signed workload container that defines how a job is executed. Cargos are reviewed, signed, and published to the Archipelag.io registry. Islands pull and run Cargos — they never execute arbitrary code. The term “Cargo” is the Archipelag.io branded name for a workload blueprint.

Examples: llm-chat (Mistral 7B), image-gen (Stable Diffusion), llm-chat-mobile (TinyLlama)

Previously known as: Workload

Karma

An Island’s reputation score. Karma reflects reliability — it goes up with successful jobs, fast responses, and uptime. It goes down with failures, timeouts, and disconnects. Higher karma means more job assignments and better placement priority.

Compute Exchange

The Archipelag.io Cargo Registry where supply meets demand. Islands set asking prices (what they want to earn per job). Consumers set bid preferences. The Coordinator matches bids to asks and clears at a fair market price.

URL: app.archipelag.io/exchange

Credits

The internal currency of Archipelag.io. Consumers spend credits to run jobs. Islands earn credits for completing them. Credits can be purchased with real money or earned through the beta program.

During beta: 1 credit = $0.01 USD

Sandbox

The security boundary around Cargo execution. Every job runs in a sandboxed environment (Docker with seccomp profiles) so that Cargos can’t access the Island’s host system, network, or other jobs. Sandbox tiers range from restricted (most locked down) to elevated (GPU/network access).

Pairing

The process of connecting an Island to your Archipelag.io account. Desktop Islands pair using a 6-character code. Mobile Islands pair by scanning a QR code. Once paired, an Island appears in your dashboard and starts receiving jobs.

Runtime

The execution engine used to run a Cargo. Supported runtimes include:

• Docker — standard container execution
• WASM — WebAssembly sandboxed execution
• llmcpp — llama.cpp-based LLM inference (used on mobile Islands)
• Browser — WebLLM/wllama in-browser execution

Island Group

A set of Islands working together to execute a single Cargo that is too large for any one Island. Island Groups are formed automatically by the coordinator when a pipeline-capable Cargo is submitted and no single Island has enough resources.

Topologies: pipeline (sequential layer sharding), expert (MoE routing), speculative (draft/verify pairs)

Lifecycle: forming → active → disbanded

Inference Ring

Another name for a pipeline-parallel Island Group. Each Island in the ring holds a subset of model layers and passes activations to the next Island in sequence. Named for the ring-like flow of data through the pipeline.

Example: A 70B model split across 4 Islands, each holding ~20 layers.

Shard Manifest

Metadata attached to a Cargo that describes how to split a model across multiple Islands. Includes the total layer count, minimum and maximum number of shards, and download URLs for each pre-split model file.

Activation

The intermediate tensor (hidden state) passed between pipeline stages in an inference ring. Activations are typically 8–16 KB per token for 7B–70B models and flow between Islands via NATS messaging.