Builder gallery

Here is the recipe.
Go build.

The architectural patterns behind every flagship on this platform — with the live demos that use them and the code to start from. These are reference patterns, not products: rough edges are documented, not hidden.

Reference pattern

Multi-head decision layer

Many classifiers, one frozen backbone: task-specific heads share a single LFM2.5-350M forward pass.

14 decision heads on one backbone22.9 MB total adapter + head weight1 forward pass per request
Reference pattern

Token-classification redaction gateway

BIO-tagged token heads that find and redact sensitive spans at the boundary — one pattern, many industry skins.

5 industry skins on one pattern<120ms measured per call (prod audit, 2026-07-07)0 documents leaving your network
Reference pattern

ColBERT late-interaction retrieval

Per-token MaxSim scoring instead of one pooled vector — retrieval that survives catalog drift.

per-token query/document interaction (MaxSim)no re-embed on catalog copy changesedge-sized runs beside the model, not a vector DB SaaS
Reference pattern

Constrained decoding

Structural validity is the decoder's job: GBNF grammars and trie logits processors make invalid output impossible, not unlikely.

100% schema-valid output by construction0 retry loops for malformed JSONgreedy temperature 0.0 for every extraction task
Reference pattern

Encoder on a frozen backbone

A small modality encoder projects structured events into a frozen LFM2.5-350M; LoRA plus task heads read out decisions.

4 internal demonstrations of the recipe~2 MB per personal/vertical adapter10–50ms warm inference, internal measurements
Reference pattern

The LEAP fine-tune loop

Task → data → train → eval → deploy: a new specialist head in minutes, gated by evals before it serves.

~90s typical 350M fine-tune wall time (build logs)5–8 regression cases minimum per capability0.0 temperature for every classification task