开发者在Colibrì项目基础上推出了Lumabri，一个探索点对点分布式大语言模型运行的实验性系统 [1]。该项目打破了传统依赖单台高性能机器的模式，允许网络中的多台普通计算机作为共享资源池，各自贡献存储和计算能力来共同运行大模型 [1]。
系统设计对Mixture-of-Experts模型特别有效，因为这类模型允许只传输小的激活信号，而无需移动庞大的专家权重参数 [1]。不过该项目目前仍处于早期阶段，正在应对网络延迟、安全验证、副本选择、故障转移和确定性执行等多个技术挑战 [1]。
Developers have introduced Lumabri, a distributed system for running large language models across multiple ordinary computers in a network rather than relying on a single machine.[1] The project reimagines how LLMs operate by treating networked computers as a shared resource pool, where each machine contributes its storage and computing capacity to form a peer-to-peer alternative comparable to Napster.[1]
The system proves particularly effective with Mixture-of-Experts models, as it can transmit only small activation signals rather than transferring massive expert weights across the network.[1] Lumabri remains in early experimental stages and is actively addressing several technical challenges, including network latency, security verification, replica selection, fault tolerance, and deterministic execution.[1] The project builds upon foundations established by the earlier Colibrì initiative.[1]
The source code is available on GitHub at https://github.com/JustVugg/lumabri.[1]