NVIDIA FLARE

license

Apache-2.0(OSI

source

public(the whole runtime, simulator, provisioning, dashboard and mobile SDK)

core features withheld

no — no enterprise tree

NVIDIA publishes the whole of FLARE under Apache-2.0, and nothing in the repository is held back for a paid tier. Its own README presents it as the open SDK platform builders use to create their own offerings, and Apheris, one of the closed platforms in this category, runs on it.

• https://raw.githubusercontent.com/NVIDIA/NVFlare/HEAD/LICENSE recorded 2026-09-27

  The LICENSE body: the verbatim Apache License, Version 2.0 text with no added clause.

• https://raw.githubusercontent.com/NVIDIA/NVFlare/HEAD/README.md recorded 2026-09-27

  README: FLARE "is a domain-agnostic, open-source, extensible Python SDK" that "enables platform developers to build a secure, privacy-preserving offering for a distributed multi-party collaboration". No paid edition, license key or enterprise tier is named.

• https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/NVIDIA%2FNVFlare recorded 2026-09-27

  Repository record for NVIDIA/NVFlare: license apache-2.0, not archived, not a fork, pushed 2026-09-19, 974 stars.

• https://ungh.cc/repos/NVIDIA/NVFlare/files/main recorded 2026-09-27

  The 5,455-file tree of NVIDIA/NVFlare at `main`: one root LICENSE, and no ee/, enterprise/, commercial/, proprietary/ or premium/ path anywhere in it. Path absence only, paired with the statement cited beside it.

PyPI downloads of `nvflare`, the package the project documents for installing both the simulator and the site runtimes. Sites run the same package, so downloads measure the product rather than a client for it.

• https://pypi.org/pypi/nvflare/json recorded 2026-09-27

  PyPI `nvflare` 2.9.0, uploaded 2026-09-04, license Apache-2.0, homepage https://github.com/NVIDIA/NVFlare.

• https://pypistats.org/api/packages/nvflare/recent recorded 2026-09-27

  last_month = 23,969 downloads for the package `nvflare`.

FLARE is the one product here that reaches all three settings at once: production sites with provisioned identities, HPC clusters through a Slurm launcher, and phones through a mobile SDK. It also publishes timing and memory figures for federated LLM training up to 72 billion parameters, which no other product in the category does.

• https://nvflare.readthedocs.io/en/main/flare_overview.html recorded 2026-09-27

  Overview: "Built-in differential privacy and homomorphic encryption" and a production path from simulation to deployment.

• https://nvflare.readthedocs.io/en/main/programming_guide/llm_fine_tuning.html recorded 2026-09-27

  The LLM fine-tuning guide: supervised fine-tuning and PEFT (LoRA) recipes that train only adapter parameters.

• https://nvflare.readthedocs.io/en/main/programming_guide/provisioning_system.html recorded 2026-09-27

  Provisioning establishes "the identities of the server, clients" and generates each site's startup kit.

• https://nvflare.readthedocs.io/en/main/user_guide/admin_guide/deployment/slurm_job_launcher.html recorded 2026-09-27

  The Slurm job launcher: FL jobs dispatched through sbatch, with squeue for live state, from a site startup kit.

• https://nvflare.readthedocs.io/en/main/user_guide/confidential_computing/index.html recorded 2026-09-27

  Confidential computing: confidential VMs on AMD SEV-SNP CPUs paired with NVIDIA H100 or Blackwell GPUs, and an Azure confidential computing deployment.

• https://nvflare.readthedocs.io/en/main/user_guide/edge_development/flare_mobile.html recorded 2026-09-27

  FLARE Mobile: an Android example app, AndroidFlareRunner and an Android DataSource interface for running federated jobs on devices.

• https://nvflare.readthedocs.io/en/main/whats_new.html recorded 2026-09-27

  What's new: "Training time and server memory across model sizes (1.7B-72B)", with "Elapsed time, 1.7B-72B (measured, 1 FL round)" and "FedAvg server peak memory, 1.7B-72B (measured, 1 FL round)".