The substrate that composes them.
Horizon is the runtime that takes the four components and runs them as one system. Cortex provides the retrieval substrate. Magnus orchestrates the cascade. Seed informs architectural choices. Soto handles modality-flexible encoding. The cascade itself, the disagreement gate, the learned stacker — all live here.
CASCADE-1, our first publicly released system on Horizon, hits SOTA on Banking77 (95.21%) and CLINC150 (97.75%). The recipe reproduces in ~500 lines of Python over public weights — the substrate is what makes it production-viable.
Horizon orchestrates the four components into one runnable system — cascading, escalation, blending, online correction.
Seed
01Compression-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
02Byte-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
03Coalition 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
04Lifelong memory.
Hash-chained immutable ledger with two-speed consolidation. 163× retention advantage over MLP baselines on continual-learning benchmarks. Online correction without retraining.
Audit. Control. Observe.
If Horizon is the substrate, Seed Foundry is the cockpit. Live audit of every prediction (via Cortex's hash-chained ledger), online correction interfaces, broker policy controls, cell-level observability, evidence packs for compliance.
Built for teams who don't just want AI in production — they want to answer for AI in production. Because “the model said so” isn't an audit trail.
Pick your point on the frontier.
Three Seed shapes for the same task. Slide between top accuracy, balanced and most efficient — and watch the size, latency and capability move together.
Beats CUD by +0.59 pp at a fraction of the size.
When accuracy is non-negotiable. The high-accuracy Seed shape clears the official Banking77 test bar above SOTA, while running on commodity CPU and fitting alongside other modules in the same memory budget.