Vertical Foundation Models
One frozen LFM2.5-350M backbone per vertical. A small modality encoder + LoRA + task heads. Four industries, one architecture: payments, motor insurance, Medicare claims, and SOC threat intelligence.
The recipe Liquid already ships in LFM2.5-Audio and LFM2.5-VL, applied to enterprise-grade structured data. Each vertical defines a small per-modality encoder that turns its native data (64 transactions, 128 telematics events, 128 Medicare claims, 64 security telemetry events) into pseudo-tokens the frozen LFM2.5-350M backbone processes natively with attention. Per-task LoRA adapters and task heads share the same backbone, so a new business objective is an adapter — not a new model build, not a new vendor procurement. The same backbone also retains its text-generation capability through zero extra parameters, letting analysts and underwriters ask natural-language questions about the very entities being scored. Four industries, four lightweight encoders, one shared 350M base.
4 specialist models
How It Works
One frozen LFM2.5-350M backbone per vertical.
A small modality encoder + per-task LoRA + task heads.
Medicare Claims: Three Payer Signals from One Model
US health payers run separate ML systems for care management, utilization forecasting, and SIU triage — each with its own actuarial vendor and refresh cadence. A single foundation model processes the full 128-event Medicare claim history with attention, surfacing admission risk, next-event forecasting, and claim integrity in under 50ms total. Adapters (~134 MB each) carry the trainable surface; the 350M backbone never moves. The same backbone answers natural-language questions about the beneficiary through a weight-tied text head — zero extra parameters. Demonstrated on CMS DE-SynPUF (public, deidentified); production deployment trains on the payer's own claims, inside the payer's VPC.
Motor Insurance: Crash Risk + Fraud + Renewal in One Pass
Motor insurers run 5-15 separate models across pricing, claims, and retention — each blind to signals that span tasks (a driver's braking pattern predicts both crash likelihood and renewal sensitivity). One PRAGMA event encoder turns raw telematics streams into backbone-native inputs. Three lightweight adapters share the same frozen LFM2.5-350M backbone and produce crash risk, fraud probability, and renewal lapse in under 50ms. A new objective — UBI tier, subrogation potential, repair fraud — is a new adapter, not a new model build. Underwriter Q&A from the same backbone, no separate NLP system.
Payments: Four Signals, One Forward Pass
Card networks maintain separate models for fraud, next-merchant, amount, and MCC enrichment — each its own training pipeline and inference cost. One transaction encoder turns 64 transactions × 15 features into 960 pseudo-tokens at d=1024, the frozen LFM2.5-350M backbone processes them with attention LoRA, and four task heads predict in parallel: fraud (BCE), next-merchant (CE over 10,003), amount bucket (CE over 16), and MCC (CE over 103). Approximately 16M trainable parameters on top of a 350M frozen base. Adding chargeback likelihood or decline-reason is a head, not a model.
SOC Threat Intelligence: Ten Signals from One Telemetry Window
Security operations centers run separate detectors for verdict, risk, attacker stage, next-tactic forecast, and reviewer signals — each its own ML pipeline and refresh cadence. One cyber telemetry encoder turns a 64-event window × 15 normalized fields into 960 pseudo-tokens, the frozen LFM2.5-350M backbone processes them with attention LoRA, and ten task heads predict in parallel: attack presence, risk, current stage, next tactic, next event type, response actions, identity compromise, lateral movement, exfil likelihood, and a benign-admin confounder. Trained on 61K real/semi-real anchors from OTRF, Splunk Attack Data, and LogHub with synthetic expansion. Adding persistence likelihood or a custom analyst label is a head, not a model.
Ready to deploy in your environment?