We built the surgical toolkit for AI knowledge transfer. Extract any capability from any foundation model and implant it precisely into any other — with zero retraining, zero GPU cost, and verified 91.7% alignment.
For years, teams had two choices when they needed a new AI capability: retrain from scratch, or distill. Both cost a fortune. Both take months. We built the third option — and it costs nothing.
Our 12-stage pipeline distils to three novel scientific contributions — each independently publishable, together forming the first complete system for cross-model knowledge transplantation.
Every concept has a precise mathematical address inside a model's weight space. We compute it in under one second using a novel gradient-decomposition technique — producing a rank-k fingerprint that uniquely identifies where and how any piece of knowledge is stored. No training required.
GPT-2 and LLaMA store the same concept in different coordinate systems. We solve the orthogonal Procrustes problem independently at each network depth — computing the exact rotation matrix between any two models' internal geometries. Residuals approach zero at all layers.
We write knowledge directly into model weights via rank-k conjugation: RTΔR — where R is the Procrustes rotation and Δ is the concept delta. Before any edit, interference detection scans for concept collisions. After surgery, an independent probe verifies the graft took at 91.7% alignment.
From loading any HuggingFace model to verified post-graft alignment — every stage built, tested, and passing. No black boxes.
Model Surgery builds on, extends, and in some areas supersedes the best existing work in neural editing and mechanistic interpretability. We are transparent about our intellectual lineage.
Meng et al., 2022. Proved facts can be located and surgically edited in transformer weights. We generalize this to arbitrary capabilities.
Read Paper →Hu et al., 2021. The adapter architecture we repurpose for concept fingerprinting — used to extract geometry, not to fine-tune.
Read Paper →Meng et al., 2022. Extended ROME to simultaneous multi-edits. We extend this concept to full capability transplantation.
Read Paper →Conneau et al., 2018. Showed that embedding spaces align across languages — directly validating our cross-model Procrustes approach.
Read Paper →"Training a 7B model costs $500,000. With Model Surgery, transplanting that capability costs $0. For the first time in history, AI capability is not a function of how much money you spent training it."
Extract French fluency from a 70B multilingual model and transplant it into a 7B English model. No bilingual data. No fine-tuning. The geometry transfers — verified.
Companies spending millions on domain-specific model training can instead surgically transplant domain knowledge. A single procedure replaces months of training expense.
For the first time: observe exactly where knowledge lives in neural networks, compare locations across architectures, verify transfers mechanistically. A microscope for AI.
From "we need this capability" to deployed in minutes, not months. Small teams now move faster than companies with 100× their budget.
Estimated annual industry savings once teams replace retraining with Model Surgery
Model Surgery is in private research beta. We are onboarding a select group of teams who want to reshape the economics of their AI development.
Patent pending. By requesting access you agree to our research terms. · research@model-surgery.com