New: CASCADE-1 — Banking77 + CLINC150 SOTA

One substrate. Four
components.

Right-sized AI for the task — from 7 MB on the edge to SOTA-class ensembles in the cloud. Auditable. Online-correctable. Multi-modal.

Read the announcement →Explore the architecture See the proof

Soto is live · 10,000 free calls / month · no credit card

95.21%

Banking77 · 5-fold strict

97.75%

CLINC150 · in-domain

~500 LOC

public components

Right-sized

4 → 1

for the task

02 — Architecture

A substrate, four components, and a control center.

Most labs ship one foundation model and tell you to fit your task to it. We take the opposite path: four task-specific research bets — each independently strong — composed into one substrate that runs the configuration your task actually needs.

Substrate — Horizon

→ Learn more

Horizon orchestrates the four components into one runnable system — cascading, escalation, blending, online correction.

Seed

Compression-first architecture search.

Find the smallest model that clears the bar. Grows from tiny (~400 params) into the right shape for your task; uses MDL loss and sleep cycles to keep the result deployment-shape, not benchmark-shape.

Soto

Byte-level encoder.

The same checkpoint runs on a $5 microcontroller or our hosted API. No tokenizer, no vocabulary to maintain, ~7 MB int8. Live today at soto.seedfrontier.ai.

Magnus

Coalition executive.

Broker, cells, disagreement gating, learned ensemble stacker, escalation policy. The “how to compose” answer that makes ensemble cascades hit SOTA with only public components.

Cortex

Lifelong memory.

Hash-chained immutable ledger with two-speed consolidation. 163× retention advantage over MLP baselines on continual-learning benchmarks. Online correction without retraining.

Control plane — Seed Foundry

Early access→ Learn more

Seed Foundry is the operator surface — audit logs, governance, evidence packs — for teams running this in production.

08 — Get started