The open engine, the shared knowledgebase. Run it on your own CPU.

The Prereasoner engine with no decoder, plus the knowledgebase everyone shares. Apache 2.0 code and open weights, small enough to run on a CPU inside your own network. Own your AI, end to end.

Try it free
View on GitHub
Open source Prereasoner running locally with reproducible results
Why us

No decoder. That's the whole trick

Encoder only, no black box

The model reads and types. It never generates, so every answer is a derivation you can inspect.

Encoder only, no black box
Deterministic and runs on CPU

Deterministic and runs on CPU

The same input always produces the same output. A frozen 0.5B core with no decoder runs on commodity CPU with no GPU required, and scales to zero when idle.

Run it on your own hardware

Apache 2.0 engine and open weights. No token bills, no cloud dependency, and no vendor that can change the model under you.

Run it on your own hardware
The shared knowledgebase

Typed world knowledge, built in the open

Grounded in real, shared sources

Geography, currencies, time zones and schema.org classes make up the typed reference layer that resolves a city to its country and currency. Free and open for everyone.

Grounded in real, shared sources
Help build the interpretability wiki

Help build the interpretability wiki

schema.org defines the classes of the world. The data that fills them is scattered and mostly closed. We are building a community wiki of open, typed reference datasets that make AI auditable. Contributors welcome at launch.

Where the line sits

Community Edition resolves the world. Structured RAG resolves your world

Community EditionStructured RAG
The engine (encoder only, no decoder)Open sourceIncluded
Open weights on Hugging FaceYesYes
Run on your own CPUYesYes
Global knowledgebase (geo, forex, schema.org)Rebuild from publisher sourcesYes
Your company's private reference tablesNot includedTenant namespace
Shared across a company's usersNot includedYes

Need to resolve gold/platinum against your own customer tier table, shared across your team? That's the tenant namespace in Structured RAG.

Deploy in minutes

Clone, deploy, done

01 CLONE

Get it

Clone the Apache 2.0 repository, then python -m engine.fetch_weights pulls the open weights. No account or token needed.

02 DEPLOY

One command

Terraform stands up one service and one Postgres in your own cloud.

03 CONNECT

Your data

Connect spreadsheets, CSV files or a database inside your network.

04 QUERY

Auditable answers

Return the result with its query and source rows.

Open source

The open engine, the shared knowledgebase. Run it on your own CPU

Use Apache 2.0 code, open weights and shared reference data on your hardware.

Try it free
Source on GitHub Weights on Hugging Face