3FS

Why we included this project

3FS is the distributed file system DeepSeek built to serve its own training and inference clusters, and it is now open source under MIT. Teams running AI workloads on hardware they own, GPU nodes with NVMe SSDs and RDMA networking, get a genuine alternative to network file systems that buckle under heavy checkpointing and dataloader traffic. It keeps the familiar file interface with strong consistency, so applications do not need rewriting around a proprietary storage API. The disaggregated architecture is what sustains multi-TiB/s read throughput across hundreds of storage nodes, which matters when training jobs move enormous intermediate datasets around. If you run your own GPU fleet rather than renting managed cloud storage, this was built for exactly the workloads that tend to bottleneck elsewhere, including checkpointing, random sample access, and KV-cache for inference.