# deepseek-ai/3FS

A high-performance distributed file system designed to address the challenges of AI training and inference workloads.

Repository: https://github.com/deepseek-ai/3FS
Canonical: https://ross.abutalabs.com/products/3fs
Language: C++
License: MIT
License Family: permissive
Topics: distributed-file-system
Last push: 2026-05-07T09:31:54+00:00

## Health v2 (maintenance only)
Score: 56/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 81, release rhythm 35, longevity 39
- inputs: {"age_days": 552, "days_push": 118, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 10166, forks 1085 (observed 2026-08-28T04:10:39.720819+00:00)

## What it is
Fire-Flyer File System (3FS) is a high-performance distributed file system from DeepSeek designed for AI training and inference workloads. It leverages NVMe SSDs and RDMA networks with a disaggregated architecture and CRAQ-based strong consistency to provide a shared POSIX-like storage layer.

## Use cases
- high-throughput shared storage for AI training clusters
- fast parallel checkpointing for large-scale model training
- random-access dataloaders without prefetching or shuffling
- KVCache backend for LLM inference
- organizing large volumes of data pipeline intermediate outputs
- running sort and analytics workloads on massive datasets

## When to choose
- you need to aggregate throughput of thousands of NVMe SSDs over RDMA/InfiniBand
- you want strong consistency with a familiar file interface for distributed AI workloads
- you need cost-effective, high-capacity KVCache for inference instead of DRAM
- you run large-scale training that requires high-throughput checkpointing

## When to avoid
- you need a general-purpose POSIX file system for small clusters or commodity networks without RDMA
- you lack dedicated NVMe storage nodes and high-speed networking hardware
- you need a mature turnkey solution rather than a system requiring careful cluster setup
- your workload is small enough for a single-node or cloud object storage solution

## Facets
- artifact type: infra-config
- maturity: active
- function: file-system, caching
- domain: microservices, machine-learning, big-data, infrastructure-as-code
- platform: cpp, self-hosted
- tags: distributed-file-system, rdma, nvme, ai-training, checkpointing, kvcache, craq, foundationdb, high-performance-storage, storage, linux

## Member repositories
- deepseek-ai/3FS (main) score 56

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:39.720819+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:19:40.218114+00:00, confidence not recorded.
  - readme: https://github.com/deepseek-ai/3FS (fetched 2026-08-28T04:10:39.720819+00:00, sha 41e2dbdadbd1)
- Data as of 2026-08-30T08:39:29.467469+00:00.