Kimi K3 2.8T 模型可在消费级硬件上本地运行

开源项目 Deltafin 实现了在普通消费级设备上本地运行完整 Kimi K3 2.8T 大模型的突破。1该项目基于 gavamedia/deltafin 分支，采用 MIT 许可证，通过从多块固态硬盘流式加载模型参数的方式，使 MacBook Pro 等设备无需云端服务即可执行大型语言模型推理。1

Deltafin 保留了 Kimi K3 的全部 16 个路由专家和完整目标模型，不进行任何参数剪枝或质量折衷。1当前项目的推理吞吐量为 0.2901 token/s（即每 token 需时 3.447 秒），相比上次更新获得了 1.9% 的性能提升。1完整模型需要 1.7 TB 磁盘空间，而流式安装模式仅需 215 GB 的初始存储空间。1该项目支持在 MacBook Pro（Apple Silicon）、Linux 和配备 NVIDIA CUDA 的设备上运行。1

该项目的核心目标是验证消费级设备能否运行前沿大模型，旨在将运行完整 Kimi K3 的基础设施成本从 Moonshot 推荐的 200 万美元大幅降低至 1.5 万美元的家庭级别配置。1

An open-source project called Deltafin allows users to run the complete, uncompressed Kimi K3 2.8T language model on consumer-grade hardware such as a MacBook Pro without any parameter pruning 1. The project streams the model from multiple SSDs, achieving approximately 0.29 tokens per second throughput 1. By storing and loading the model from disk rather than relying on expensive server infrastructure, Deltafin aims to demonstrate that cutting-edge large language models can operate on standard personal computers.

The Deltafin fork, based on the gavamedia/deltafin repository under MIT license, maintains all 16 routing experts and the full target model without compromising quality 1. The current setup requires 1.7 terabytes of disk space for the complete model, though a streaming installation mode needs only 215 gigabytes of initial space 1. The project supports Apple Silicon MacBook Pro, Linux, and NVIDIA CUDA platforms 1. According to the developers, their goal is to enable a fully functional Kimi K3 deployment on a home setup costing approximately $15,000, contrasting sharply with the $2 million in infrastructure recommended by Moonshot 1.