Skip to main content
AI Foundry puts open-source biological models behind one uniform REST surface — /v1/foundry/models/…. Rafflesia owns the public resource, content-addressed result storage, metering, and provenance; heavy inference runs on external GPUs. Every result is addressed by object id and every call is recorded against an explicit model release.

​
Workloads

Browse the catalog with GET /v1/foundry/models (filter by ?workload=), then call the endpoint for that workload:

Structure prediction

POST …/predictions — durable jobs. Boltz-2, Chai-1, Protenix, AlphaFold2, RoseTTAFold3, ESMFold2.

Embeddings

POST …/embeddings — sync. Protein (ESM C, ESM-2), DNA (Evo 2, HyenaDNA, Caduceus), RNA (RNA-FM), and 3Di (SaProt, ProstT5).

Protein design

POST …/designs — sync. Sample sequences on a fixed backbone with ProteinMPNN and LigandMPNN.

Variant effects

POST …/variant-effects — sync. Substitution log-likelihood ratios from the ESM-1v ensemble.

​
How a call works

  • Structure prediction is a durable job. POST …/predictions returns a job_id; you poll it, and the result surfaces as a content-addressed structure object. Everything else — embeddings, designs, variant effects — returns inline.
  • Results are objects. Structures in and out are ObjectRefs (object_id, uri, sha256), so downstream work is reproducible.
  • Calls are safe to retry. Every POST accepts an Idempotency-Key.
  • Releases are explicit. Each call names a model_release, and the provider must echo the exact release that produced the result.