Core-1 is a language model designed to run on ordinary CPUs instead of rented GPUs. Two ideas, stacked.
Every weight is a 32 bit number, and every step of inference multiplies them. That is why models need big GPUs.
Every weight is forced to −1, 0 or +1 during training. Packed at two bits each, the model gets sixteen times smaller.
Zero means skip. Plus one means add. Minus one means subtract. Plain integer math that cheap processors are good at.
Attention keeps a cache that grows with every token. Core-1 uses a state space backbone with a fixed size state instead.
Status: proof of concept. A small ternary state space language model that learns and packs sixteen times smaller, validated at toy scale.
A small coding model trained for secure code review. Point it at source code and it flags vulnerabilities and suggests safer implementations, fully offline.
Strongest on SQL and command injection. The model card lists where it still falls short.
Selective about projects. Ambitious about the ones I take: models, engines, and the tools around them.
Start a project →benedictpatrickjohn@gmail.com