FERMI
Nuclear-specific sentence embedding models, open source on Hugging Face.
FERMI is the retrieval layer that powers Neutron’s search. Trained on 53+ million pages of NRC data using Oak Ridge National Laboratory’s Frontier supercomputer, FERMI understands nuclear language at a depth that general-purpose AI can’t reach. It’s the reason Neutron is 40% more accurate on nuclear-specific tasks.
The retrieval layer, not the language model
FERMI is a sentence embedding model, the AI that powers search and retrieval. It is not a large language model like ChatGPT or Claude.
Here’s the difference that matters: when you ask a question, the LLM generates the answer. Without retrieval, it generates that answer from training memory, a fuzzy recollection of whatever it has read. In nuclear, answers can’t rely on memory. They have to be grounded in the actual records and documentation: the procedures, the rules, the regulations, in the correct revision. Between one revision and the next, a procedure or regulation can change in ways small or sweeping, and the difference matters.
FERMI is what finds the right documents and records and gives the LLM the accurate, full context. Without accurate retrieval, the best LLM in the world is guessing.
General embedding models are trained on internet content. In nuclear, that’s a problem.
Sees “trip” and thinks vacation. Beach bookings, travel deals.
Sees “trip” and retrieves emergency operating procedures, safety system documentation, and Technical Specification requirements.
The difference is 53+ million pages of nuclear-specific training data instead of the open internet.
What makes nuclear retrieval different
FERMI handles the hard part: getting the right nuclear-specific information to the LLM. Neutron pairs FERMI with frontier models from OpenAI, Anthropic, Google, NVIDIA, and others. You bring the model; FERMI makes sure it gets the right information.
Training data
53+ million pages of NRC ADAMS records: license amendment requests, inspection reports, Safety Evaluation Reports, Licensee Event Reports, and technical correspondence from across the U.S. operating fleet. Trained on ORNL’s Frontier supercomputer through a collaboration with Oak Ridge National Laboratory and the U.S. Department of Energy.
Custom tokenizer
General-purpose tokenizers split nuclear-specific terms into meaningless fragments. A term like “NRC” might get broken into pieces the model can’t interpret as a single concept. FERMI uses a custom tokenizer built for nuclear language, keeping industry terminology intact.
Temporal awareness
The right answer in nuclear often depends on the right era of documentation. NRC and IAEA standards have been revised across decades, and each plant has its own licensing basis vintage. Getting the right answer means retrieving the right-era regulation matched to the right-vintage plant basis. FERMI and Neutron are designed with this temporal dimension in mind.
Open models, open benchmarks
FERMI models are publicly available on Hugging Face.
FermiBench
FermiBench is our open retrieval benchmark for nuclear documentation: 156 queries with ground-truth relevance labels across 4,558 NRC ADAMS documents. Published openly so anyone can reproduce FERMI’s retrieval scores or evaluate their own models on the same nuclear data.
View on Hugging Face →NuclearBench
Developed with Idaho National Laboratory, NuclearBench is the first AI benchmark suite for nuclear applications. Open source, so any utility, regulator, or vendor can evaluate AI performance on nuclear data.
Coming soonBuilt with leading nuclear research institutions
Oak Ridge National Laboratory
FERMI was developed in collaboration with ORNL and the U.S. Department of Energy, trained on ORNL’s Frontier supercomputer.
Idaho National Laboratory
NuclearBench, the first AI benchmark suite for nuclear applications, was developed in collaboration with INL.
See FERMI in action
Talk to our team about nuclear-specific retrieval for your AI stack.