Newsroom
Eight Billion Parameters in One Gigabyte
We're adding sub-2-bit LLM inference to our macOS Islands. An 8B model now fits in 1.1 GB on disk, 2.4 GB at runtime, and runs at 66 tokens/sec on a MacBook M4 — through Archipelag.io's actual production code path, not a vendored CLI. Here's why it matters and what we had to fix to get there.
The Math Says We Cannot Read Your Prompt
Most cloud AI runs on a contractual privacy promise: we say we won't read your prompts. We're moving to a technical one: we cannot. Here's what we shipped on Apple Silicon, what it actually protects against, and what it doesn't.
Europe Without Datacenters
AI inference is here to stay. Europe is weak on the infrastructure that runs it and strong on the values that should shape it. Datacenter-less inference is how the two reconcile.
Your Phone Is an Island
We turned an iPhone into a compute node. A prompt typed in a browser, answered by a language model running on someone's phone across town. Here's how it works, and why it matters.
Archipelag.io Enters Open Beta
We're opening Archipelag.io to everyone for three months. Come use AI, lend your GPU, and help us figure out what works.
Introducing Archipelag.io
We built a distributed compute network that routes AI Cargos to nearby hardware instead of distant data centers. Here's the story of why, and what it actually does.