Platform · Module
Model Hub
The only path to any model, routed by task, data class, quality and cost
Model Hub is the single gateway between One3 and every AI model. It handles registry, routing, data-classification rules, fallbacks, caching, bring-your-own keys and a cost record for every call.
One gateway
No One3 code calls a model provider directly. Every call goes through Model Hub with a task class, data classification, budget, trace and cost. Swapping a model becomes a configuration change plus an evaluation run, not a rewrite.
Registry and routing
The model registry records each model's provider, licence (linked to the Fact Register), region, modalities, context size, price reference and status. Routing works by task class: classify, extract, summarise, draft, reason, code, plan, vision, embed, rerank, speech recognition, text-to-speech, image and video.
Data stays where it should
From Wave 1, data-classification routing sends restricted and sensitive personal data only to in-region or local models, with a masking policy. Fallbacks, retries and a circuit breaker handle provider errors and timeouts.
Quality and cost together
Planned for Wave 2:
- Confidence cascade: the cheapest qualifying model goes first, escalating on schema failure, self-check or a confidence threshold.
- Eval-driven routing: route tables change only when the evaluation harness shows equal or better quality, or equal quality at lower cost.
- Per-tenant semantic cache that never crosses tenants, plus provider prompt caching where supported.
- Bring-your-own provider keys per tenant, still traced, with model cost not charged to credits.
- Licence and terms gate: a model is not routable until its commercial-use and output terms are recorded in the Fact Register.
- India-hosted routing options as a tenant policy (provider availability to verify).
- An admin view of allowed providers and models, residency policy, BYOK status and cost by task class.
Per-tenant LoRA adapters, One3's own Action and extraction models, and local-only sovereign routing for One3 Box are planned for Waves 3–5.
Features
What Model Hub includes
Gateway with cost row per call
LiteLLM proxy with local and frontier providers and per-tenant keys. W0.
Model registry
Provider, licence, region, modalities, context size, price reference and status. W0.
Task-class routing
Fourteen task classes from classify to video. W0.
Data-classification routing
Restricted and sensitive data stays in-region or local, with masking. W1.
Fallbacks and circuit breaker
Retries and failover on provider errors and timeouts. W1.
Per-tenant semantic cache
Never shared across tenants; provider prompt caching where supported. W2.
Confidence cascade
Cheapest qualifying model first, escalating when needed. W2.
Eval-driven routing
Routes change only on equal-or-better evaluated quality or lower cost. W2.
Bring your own keys
Tenant provider keys in the secret manager, still traced. W2.
Budgets and priority classes
Interactive ahead of approvals ahead of batch; batch queue for non-urgent jobs. W2.
Licence and terms gate
Models routable only after terms are recorded and verified. W2.
India-hosted routing
In-region providers or One3-hosted open-weight models as a policy (to verify). W2.
Tenant admin view
Allowed models, residency, BYOK status and cost by task class. W2.
Workflow
Route a sensitive extraction
- 1
An app requests an extraction on a document labelled sensitive_pii.
- 2
Model Hub restricts candidates to in-region or local models.
- 3
The cheapest qualifying model runs first.
- 4
Schema failure escalates to a stronger model.
- 5
Cost and trace are recorded against the tenant and Action.
AI
AI skills and Digital Employees
Router
Chooses a model by task class, classification, eval table, budget and cascade. Always on.
Eval-driven route promotion
Canary and rollback for route changes, human-approved. W2.
Semantic and prompt caching
Reduces credit use where eligible. W2.
Digital Employees that use this module
- AI Agent Governance Auditor
They act through the same governed Actions as people; risky steps wait for approval.
Frequently asked questions
Which models does One3 use?
Model Hub is model-neutral. It routes between local and frontier providers based on task, data classification, evaluated quality and cost. The registry records each model's licence and region.
Will my restricted data go to a foreign provider?
Data-classification routing, planned for Wave 1, sends restricted and sensitive personal data only to in-region or local models.
Can I bring my own OpenAI or other provider key?
Bring-your-own provider keys are planned for Wave 2 on higher plans.
See Model Hub in your workflow
Tell us about your team and the outcomes you want. We will walk you through the apps, the Digital Employees and the plan for your industry.