# Reame

A CPU-first LLM inference server built around ternary weight models, running an 8-billion-parameter model in about 2.2 GB of memory on a free-tier ARM cloud instance.

Reame is an LLM inference server designed for ordinary CPUs, built specifically around ternary models, whose weights are limited to negative one, zero, or positive one. Because a ternary model's matrix products reduce to efficient integer operations, an 8-billion-parameter model can run in about 2.2 GB of memory, small enough for a free-tier ARM cloud instance, with no GPU required and no per-token billing.

It combines two engines: nucleo, a custom engine built from scratch that converts a model once into its own interleaved 2.125-bit format and handles ternary weight computation, and llama.cpp for everything else. The project publishes measured decode and prompt-processing throughput against the model vendor's own llama.cpp fork on the same ARM hardware, along with the full benchmark methodology, including cases where its own results are worse.

It exposes a single OpenAI-compatible API that any existing client can connect to, is written in C++, and is MIT licensed.

Website: https://swellweb.github.io/reame/