Skip to content

Frontier Models

The factory builds on external models that no single box could hold alone — in two distinct roles: the reasoning layer that interprets and orchestrates, and the domain-science models that generate molecules and fold structures. Each connects on demand and registers as a native capability with an honest status, so nothing here pretends to run locally, and nothing claims more than it delivers.

The reasoning layer

  • Anthropic (Claude) — the reasoning LLM behind the Precision Intelligence Engine and all eight intelligence agents. The "intelligence" in intelligence agent. Reached as a hosted API — the specific model is configurable, and the layer degrades honestly (never mock-served) when a key is absent.

Domain-science models

Heavy or partner-gated models — too large for one box, or available only through a partner — reached on demand via elastic burst to a remote GPU:

  • NVIDIA BioNeMo — GPU molecular microservices: MolMIM & DiffDock (live) and GenMol (planned) for the Therapeutic Discovery Engine.
  • Chai Discovery — Chai-1 (complex co-folding · planned) and Chai-2 (de-novo binder & antibody design · gated): predict a molecular complex, or design a brand-new binder — and chain the two (design → validate).

Also on the roadmap

Registered and connected the same way — heavy or containerized, honest status, elastic burst:

  • RFdiffusion   planned — de-novo protein backbone design (→ ProteinMPNN → Chai-1).
  • Evo 2   planned — genome-scale foundation model.

Honest by construction

The statuses here are the registry's — see the live Capability Maturity Matrix. The domain-science models burst to a remote GPU (elastic burst), never pretending to run on the box; the reasoning layer is a hosted API — configurable, and never mock-served. Any clinical use is decision support for a qualified clinician, never autonomous diagnosis.