PolyCode

An experimental coding model that denoises a whole syntax tree at once, rather than predicting a program one token at a time.

Overview

PolyCode is one of two model programs that continue after the frontier wind-down. It is small by design: the goal is a coding model that runs usefully on hardware an engineering team actually controls, rather than one that needs a datacenter behind it.

It is an experimental model and is presented as one. It is not a drop-in replacement for a large general-purpose coding assistant.

How it differs

Most code models predict the next token, which means a program is correct only if every step along the way happened to be correct. PolyCode instead starts from a noised structure and denoises it toward a valid syntax tree, so the output is structurally sound by construction.

The practical effect is that it fails differently. When PolyCode is wrong, it tends to be wrong about intent rather than producing something that will not parse.

Where it runs

PolyCode targets local and air-gapped installations, the same environments the Xora family was built for. Nothing leaves the machine it runs on.

Ongoing work is published through Infinity Intelligence Labs.

Specification

FamilyPolyCode
ApproachDiffusion over syntax trees
StatusActive development
TargetLocal and edge deployment
WeightsClosed
AvailabilityLabs research and selected deployments