Foundry.
Try the live demo
The on-premise AI platform

Foundry.

From data to deployment, the definitive AI workflow.

Run it all on-premise, inside your own infrastructure.

Try the live demoTalk to us

Self-hosted · Hugging Face compatible · your infrastructure

Verified, not vague

Built on facts you can check.

Every number is counted from the source code — not estimated.
Repository typesmodels · datasets · spacessrc/lib/config/hub.ts
Distributed-training frameworksDDP · Ray · Horovod · DeepSpeedsrc/lib/reactors/distributed-training.ts
HFHugging Face compatiblepoint HF_ENDPOINT and gosrc/lib/api/hf-response-formatter.ts
LakeFSVersioned storagebranch · commit · tag, any sizesrc/lib/lakefs/client.ts
Hugging Face compatible

Already using Hugging Face? Just point HF_ENDPOINT here.

Your existing huggingface_hub code keeps working against Foundry for core workflows — upload, download, and commits.

# Point the official client at Foundry
export HF_ENDPOINT=https://your-foundry.example.com

from huggingface_hub import HfApi, hf_hub_download

api = HfApi(token="hf_...")
api.upload_file(
    path_or_fileobj="model.safetensors",
    path_in_repo="model.safetensors",
    repo_id="your-org/your-model",
)

config = hf_hub_download("your-org/your-model", "config.json")

Core repository workflows are drop-in compatible. Some advanced Hub endpoints are still in progress.

  1. huggingface_hub

    official client

  2. HF_ENDPOINT → Foundry

    HF-compatible API · _foundry_* fields

  3. LakeFS

    versioned storage

  4. S3-compatible

    MinIO / Wasabi

End to end

One platform. The whole AI lifecycle.

From raw data to a running deployment, without stitching separate tools together.

  1. 01
    Data

    Bring your files, documents and URLs.

  2. 02
    Data Forge

    Turn raw sources into training data.

  3. 03
    Hub

    Version models and datasets on LakeFS.

  4. 04
    Reactors

    Train and run on multi-cloud GPUs.

  5. 05
    Deploy

    Ship models and spaces.

Built for the whole team

Everything your AI work needs.

01Repository Hub

Models, datasets and spaces — versioned like code.

Host all three repository types on a Git-style backend powered by LakeFS. Branch, tag, commit and open pull requests on data of any size.

  • Branches, tags and commit history on LakeFS
  • Pull requests with cherry-pick merges
  • Discussions, @mentions and likes
  • Monaco code viewer, editor and diff
  • Data Studio: in-browser SQL on Parquet, CSV and JSON via DuckDB
src/lib/lakefs/client.ts
models/acme/llm-7b · main
  1. fix configpr-42
  2. update weights
  3. add tokenizerv1.0
3 branches2 tagsversioned
02Data Forge

From raw documents to fine-tuning data.

A six-step pipeline that ingests your sources and generates high-quality training data with an LLM of your choice.

  • Six steps: Upload, Ingest, Create, Curate, Export, Publish
  • QA, Chain-of-Thought, Summary and multimodal generation
  • Parsers for PDF, DOCX, PPTX, HTML, images, URLs and YouTube
  • Quality-score curation for QA pairs
  • Export to JSONL, Alpaca, ChatML or fine-tuning formats
foundry-dataforge/
data forge · session
Upload
Ingest
Create
Curate
Export
Publish
quality ≥ 7.0
8.4kept7.1kept6.2filtered out
export
JSONLAlpacaChatML
03Reactors

Multi-cloud GPU, orchestrated.

Launch clusters, jobs and services across clouds with SkyPilot — from a single console.

  • Cluster, job and service lifecycle on SkyPilot
  • Distributed training: PyTorch DDP, Ray, Horovod, DeepSpeed
  • GPU recommendations with estimated cost and spot savings
  • Web terminal and live logs
  • Autostop presets and organization quota policies
foundry-reactor/
cluster · a100 · upup
GPUs
est. cost
queuecontrol2jobs5io10
dataset · train-00000.parquet
SELECT prompt, label FROM dataset LIMIT 5
Illustrative query result
promptlabel
Summarize the passage.summary
Translate to French.translation
Is this review positive?positive
Extract the date.entity
Explore your data

Query datasets right in the browser.

DuckDB-powered SQL over Parquet, CSV and JSON — no download required.

  • In-browser SQL with DuckDB WASM
  • Parquet, CSV, JSON and Arrow (plus .gz)
  • Image-text pair gallery viewer
  • Safetensors tensor metadata viewer
  • Auto-loads split files like train-*.parquet
src/components/repos/data-studio
Built to collaborate

Review, discuss, and merge — together.

Pull requests, discussions and mentions live right next to your models and data.

  • Pull requests with cherry-pick merges
  • Threaded discussions with @mentions
  • Likes and follows across repos and people
  • Organizations with owner, admin and member roles
  • In-app and email notifications
src/lib/api/pr-handlers.ts
Pull request #42merged
maintainer

@reviewer can you take a look at the eval results before we ship?

reviewer

Looks good — metrics held up. Merging now.

3likes
space · gradio-apprunning
replicas
2
region
us-east-1
service · my-modelPreview
replicas
1
up
Deploy anywhere

Runs on your infrastructure

Install the full stack on-premise with Docker Compose, on your own servers and GPUs. The core stack runs inside your private network; your data stays with you.

  • Spaces — compute-backed apps
  • Model services on ReactorsPreview
  • Distributed training: DDP, Ray, Horovod, DeepSpeed
  • Autostop and org quotas for cost control
foundry-reactor/
Enterprise ready

Security built into the code, not bolted on.

Compliance-friendly by design, with defense in depth across the stack.

config.env · scanning
1REGION=us-east-1
2API_KEY=secret value redacted
3DEBUG=false
secret detected · uploader notified
src/lib/security/secret-patterns.ts

Token rotation

JWT refresh-token rotation with reuse-attack detection.

src/lib/auth/token-service.ts

2FA with replay defense

TOTP, AES-256-GCM encrypted, with replay protection.

src/lib/auth/totp.ts

Content scanning

Secret detection patterns and ClamAV malware scans on uploads.

src/lib/security/secret-patterns.ts

PII masking

Emails, IPs and user IDs masked in production logs.

src/lib/logging.ts

Roles and permissions

Three-tier org roles mapped to repo read, write and owner.

src/lib/config/organization.ts

CSRF and scoped tokens

Double-submit CSRF and scoped personal access tokens.

src/lib/auth (CSRF)
Govern with confidence

Access and audit, by design.

Role-based access, audit trails, and content moderation, built into the platform.

  • Org roles map to repo read, write and owner automatically
  • Audit logs for sensitive operations
  • Content moderation — secret and malware scans alert admins
  • Read-only admin impersonation for support
src/lib/config/organization.ts
Role permission matrix: roles mapped to view, contribute, manage and owner capabilities
RoleViewContributeManageOwner
Member
Admin
Owner
Developer experience

Build with the Python SDK and CLI.

A clean SDK and a rich command-line client for everything from upload to download.

Read the docs
from foundry_hub import FoundryApi, upload_file

with FoundryApi(token="...") as api:
    models = api.list_models()
    info = api.model_info("your-org/your-model")

# Large files use multipart upload automatically
upload_file("your-org/your-model", "./model.safetensors")
Before you adopt

Questions teams ask first.

Straight answers, grounded in what the code actually does.

Yes. Set HF_ENDPOINT to your Foundry server and the huggingface_hub library keeps working for core workflows — upload, download, and commits. Some advanced Hub endpoints are still in progress.

Deploy Foundry inside your infrastructure

On-premise for governments, public institutions, and enterprises. Explore the live demo, then talk to us about deployment.

Try the live demoTalk to us
Foundry — From data to AI deployment