02Primitive

LRM v2

Lorem ipsum dolor sit amet — a novel transformer architecture designed from the ground up to be populated by Knowledge Modules at runtime.

Novel architectureIn active build
Overview

A reasoning core that borrows its knowledge.

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris.

Nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.

Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum. Sed ut perspiciatis unde omnis iste natus error sit voluptatem.

Key innovations

What makes it patentable.

Claim 01

Split-layer topology

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Reasoning layers and knowledge-bank layers trained as separable substrates.

Claim 02

Empty-bank pretraining

Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua — knowledge-bank layers begin blank and are populated at inference.

Claim 03

KM hot-swap runtime

Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.

Claim 04

Reasoning-only loss

Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur during pretraining.

Claim 05

Composable capability stack

Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.

Claim 06

Serving-cost collapse

Sed ut perspiciatis unde omnis iste natus error sit voluptatem — small core plus loaded KMs approaches frontier performance.

How it works

The method, in steps.

Step 01

Pretrain the core

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Reasoning layers train on language, logic, math, and code with knowledge-bank layers held empty.

Step 02

Write the modules

Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua — capability content is written into portable KMs via DeltaWrite.

Step 03

Load at runtime

Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat at inference time.

Step 04

Compose & swap

Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore across domains without retraining the core.

The numbers

Measured, not claimed.

Core parameters
7B

Lorem ipsum — reasoning-only substrate.

Bank capacity
64 KMs

Dolor sit amet loaded concurrently per inference.

Frontier parity
~92%

Consectetur adipiscing elit on MMLU-Pro with loaded KMs.

Serving cost
0.18×

Adipiscing elit vs. comparable frontier dense model.

Applications

Where it deploys.

Composable foundation serving

Lorem ipsum dolor sit amet, consectetur adipiscing elit — a single reasoning core serving many verticals via hot-swapped KMs.

Cost-efficient frontier inference

Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua at a fraction of current serving economics.

Sovereign-AI deployments

Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris — core plus jurisdiction-specific KMs for regulated markets.

Edge inference

Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore on consumer hardware with targeted KMs.

Engage

License, replicate, or co-develop.

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Licensing, training-partner, and co-development conversations are open for LRM v2. Technical deep-dives available under NDA.

Read the blog post