About ModelForgeLab

ModelForgeLab is an open adapter registry built for teams that train and ship fine-tuned models. We started in 2024 after spending too much time wrangling ad-hoc file shares and undocumented weight dumps — the ML community needed a proper versioned registry the way software needs package managers.

What we build

The core of ModelForgeLab is a version-controlled store for LoRA adapters and fine-tuned checkpoints. Every artifact gets a sha256 digest, a semantic version, and format-aware storage that knows the difference between a safetensors file and a GGUF quantisation. On top of that we layered a fine-tune queue, a streaming playground, and a CLI that makes pulling a remote adapter as simple as pulling a container image.

The entire stack — registry, queue, and API — can run on a single modest VPS or be self-hosted inside your own infrastructure. No adapter weights, datasets, or credentials ever touch a third-party service unless you explicitly configure one.

Our principles

• Privacy by default. We collect nothing we do not need. The auth flow uses a signed cookie with an attempt counter; no email, no password, no PII is persisted anywhere.
• Open core. The community edition is fully functional with no feature gates. Enterprise features (dedicated support, managed hosting, SSO) are addons, not unlocks.
• Verifiable artifacts. Every download is verified against a sha256 checksum before it is written to the local cache. No silent corruption.
• Self-hostable first. We design for the case where you run this yourself. Managed hosting is a convenience, not a requirement.

Get in touch

Questions, partnership enquiries, or feedback — write to support@modelforgelab.cloud. For bug reports and feature requests, open an issue on GitHub.