SuperWVY

A new architecture, not a new finetune. SuperWVY is our language-native architecture research: building model structure around how language is actually composed — letters, morphemes, words, phrases, and higher-order meaning — instead of treating text as only a flat stream of prediction tokens.

Why language-native

Most language models compress text into token pieces and learn the next prediction from scale. SuperWVY explores a different route: make the architecture aware of the hierarchy humans already use when they read and reason. Letters form morphemes, morphemes form words, words form phrases, and phrases carry meaning inside a larger context.

• Language hierarchy — structure across letters, morphemes, words, phrases, and meaning units
• Neural learning — pattern recognition, conversation, and flexible understanding
• Symbolic structure — explicit scaffolding the model can use while reasoning
• The bridge — training methods that connect learned patterns with language structure

Reasoning goal

The aim is not to make the biggest model possible. The aim is to make a smaller model that can use information more precisely: read context, organize it, challenge weak assumptions, and solve the problem in front of it. That means cleaner datasets, stronger intermediate reasoning, and architecture choices that help the model build with information instead of jumping straight to a memorized answer.

Status

Design documents

The deeper specs are still being prepared for public release. For now, this page states the direction without exposing implementation details that are still moving.

# SuperWVY public notes
# language hierarchy
# symbolic / neural bridge
# efficient reasoning training

Related: Savvy applies the same reasoning-first philosophy to autonomous research.