ARIA Protocol is a peer-to-peer (P2P) protocol based on 1-bit quantized models (Microsoft's BitNet b1.58 research) specifically designed for distributed AI inference. The ternary weights (-1, 0, +1) transform matrix multiplication into addition and subtraction operations, making it possible to run inference on regular CPUs without requiring GPUs.

A community member conducted benchmark tests on an AMD Ryzen 9 7845HX (DDR5-5600, 8 threads) and an Intel i7-11370H Tiger Lake (DDR4, 8 threads). The results showed that the AMD achieved 89.65 tokens per second on the 0.7B model, while the Intel surpassed it with 77 tokens per second on the larger 2.4B model. This advantage may be attributed to Tiger Lake's native 512-bit AVX-512 instruction set, which optimizes ternary operations.

In terms of energy consumption, the protocol requires between 11 to 66 millijoules per token, which is 70-82% lower than GPU inference for equivalent models. The protocol supports P2P networking, pipeline parallelism (distributing model layers across different nodes), a ledger to record each inference for traceability, and an OpenAI-compatible API.

Regarding memory architecture, ARIA Protocol implements a comprehensive cognitive memory model based on psychological literature, including episodic memory, semantic memory, procedural memory, working memory, and prospective memory. Prospective memory stands out as a key feature, enabling functionality like "remember to mention this when we discuss topic X next time."

The project is licensed under MIT, built on Python 3.10+, and has 196 passing tests. The desktop application supports Windows, macOS, and Linux.

GitHub repository: https://github.com/spmfrance-cloud/aria-protocol

Project website: https://spmfrance-cloud.github.io/aria-protocol