Model Surgeon is an operating theater for neural networks. Instead of treating AI models as sealed black boxes, it opens them: load a real model file in your browser and see its complete anatomy rendered in interactive 3-D — every layer, every tensor, every connection. From there you operate. Cut out dead weight, graft in components, merge two models with precision-controlled blending, halve precision for edge deployment — all of it visual, all of it reversible until you commit, and all of it recorded in a surgery log that captures exactly what was done. Critically, your model is never executed during any of this: files are parsed as pure data, which means even a hostile file off the internet has nothing to run, and your weights never leave your session.
What separates Model Surgeon from viewers and scripts is that every cut is measured and every result is provable. Built-in diagnostics show you the consequence of each operation — output drift, size and compute savings, forensic scans for tampering and hidden payloads, provenance analysis that can tell you whether one model was derived from another even after renaming. When you export, your model ships with a cryptographically signed record of its entire surgical history, sealed to the file's own fingerprints. Anyone you send it to — a client, an auditor, a teammate — can verify that record offline, for free, forever, with an open-source verifier whose full specification is public: change a single byte anywhere in the package and the verdict flips from VERIFIED to TAMPERED. Trust isn't a claim in a README; it's a fifteen-second check anyone can run.
The result is a complete loop — open, cut, measure, prove — that previously required a stack of disconnected tools and a lot of unrecorded tribal knowledge. Surgeries export as replayable recipes: human-readable files that apply the same operations across entire fleets of models, with a growing community registry of shared, validated recipes to start from. Whether you're an engineer shrinking a model for deployment, a security team vetting third-party weights, or a student seeing inside a transformer for the first time, the promise is the same: real surgery on real models, with receipts. Model Surgeon, from FALCONS.AI — surgery, not slides.
FALCONS.AI is an AI company based in Delaware, USA and trusted globally — the team behind the world's most-downloaded NSFW detection model, with over 1.3 billion downloads on Hugging Face protecting platforms and their users, including children, from harmful content. The company builds AI that is efficient, specialized, and safe: visual content safety served in real time via cloud API or fully on-premise for data sovereignty, adversarial testing instruments for image models, and published research held to the standard of evidence. Founded by Michael Stattelman, Bobbi Stattelman (Chief Operating Officer), and Nizamuddin Mohamed (Chief Technology Officer), FALCONS.AI operates on one conviction across everything it ships: trust in AI should never have to be blind. Model Surgeon is that conviction made into an instrument.
More at falcons.ai · falcons.ai/company