Built on NVIDIA Nemotron
Safety comes from the architecture.
Judgement comes from Nemotron. Pick the right size for each job, and run it where the data lives.
Every AI in RecoveryPact is a Nemotron model: the two negotiating guardians, the auditor, the guard that screens what they read, and the embeddings behind their memory. We benchmarked the same eight incidents across the Nemotron family, with and without our reasoning scaffold, and with a mixed-size team.
The right Nemotron for each role
Every Nemotron size
The scaffold makes a 3B-active Nemotron dependable
Big model where judgement matters, small model everywhere else
Nemotron screens what the agents read
Incident data should not leave the building
A recovery agent reads exactly what companies guard most: which sessions touch payments, which customers are failing, whether the financial close is running. With open Nemotron weights the whole pact can run on-premises. NVIDIA rates DGX Spark for local inference on models up to about 200 billion parameters, which covers the Lightning and Super classes we benchmarked here. Finding 3 shows how much of the work a small model can carry.
Our runs used the hosted NVIDIA NIM API. Running on a DGX Spark is the next step, not something we have measured.