Work at the floor of arithmetic.
Fermion Research trains foundation models natively in a ternary-family weight format and builds the engine that serves them. The methods that hold frontier capability at this depth are developed here and are unpublished. The teams are small and own whole systems.
Open roles
Member of Technical Staff, Inference Systems
Own single-stream decode across four backends: CUDA, Metal, ARM, and x86. One artifact, one speed discipline, kernel by kernel.
Member of Technical Staff, Training
Train models natively in the format they ship in. Closing the gap between fp16 pretraining and native-format capability is the lab's core research problem.
Member of Technical Staff, Evaluation
Build the infrastructure behind every published number: the accuracy batteries, the decode-rate rigs, and the token-for-token verification that gates each release.
The whole stack, weight format to engine, is built here.
Apply
Send what you have built to careers@fermionresearch.com