One3 Ecosystem · Module
Model Marketplace
Licence-checked models and adapters, tested on your own tasks before use
Model Marketplace will let tenants choose, evaluate and enable licence-checked models, adapters and later datasets as governed routes in Model Hub. No model is enabled until it has been evaluated on the tenant's own golden tasks.
Model Marketplace is planned for W5 (Years 4–5) and is sold on revenue share, metered by tokens and compute hours. Before W5, Model Hub offers only first-party routes.
The problem
Model hubs publish model cards and generic benchmarks. They do not tell you how a model performs on your own workflows and Actions, or whether its licence fits your deployment.
How it works
Publisher side. A publisher submits a model or adapter with a model card and licence. Artifact checks require safe serialisation formats (no executable pickle payloads), malware scanning, hashes, signatures and provenance. The licence record is reviewed for commercial use, redistribution and deployment eligibility (cloud, One3 Box, air-gapped, edge). Safety evaluations test jailbreaks, prompt-injection susceptibility, PII leakage and harmful output in English and launch Indian languages. Critical failures block the listing.
Tenant side.
- View the card, licence terms, publisher-reported benchmarks (labelled as such) and a cost and latency estimate.
- A tenant admin accepts the licence, which is recorded.
- Run the tenant's own golden tasks and skill eval suites against the candidate in a sandbox with redacted data, and compare quality, safety, cost and latency with the current route.
- If it meets thresholds, a governance check runs for high-risk uses.
- Enable as a governed route with allowed data classes, allowed skills and agents, region, fallback route and budget.
- Canary with rollback thresholds, then promote and monitor drift against the eval baseline.
Rules
Adapters trained on any tenant's data are never publishable. Hosting can be One3 Compute, a provider with declared residency, or One3 Box where the licence permits. Bring-your-own provider contracts and keys are supported for listed routes. A kill switch can disable a listing or route across all tenants, with fallback.
Features
What Model Marketplace includes
Listing types
Open-weight models, provider-hosted routes, LoRA adapters, embedding and reranker models, and speech models.
Model cards
Intended and out-of-scope uses, languages, training-data summary, limitations, size, context length, hardware class and eval results.
Licence term records
Commercial use, redistribution, on-prem and One3 Box eligibility, attribution and use restrictions, checked against the tenant's deployment profile.
Gated access and licence acceptance
Tenant admin acceptance is recorded, with publisher approval for gated models.
Artifact security
Safe serialisation formats only, malware scan, hashes, signatures and provenance.
Benchmarks and business eval suites
Publisher-reported benchmarks labelled as such, plus One3-run synthetic business eval suites.
Tenant-specific evaluation
Run the tenant's golden tasks and skill evals in a sandbox and compare quality, safety, cost and latency with the current route.
Enablement as governed route
Enable in Model Hub with allowed data classes, skills, region, fallback route, budget and canary percentage.
Safety evaluation
Jailbreak, prompt injection, PII leakage and harmful output tests in English and launch Indian languages.
Hosting options
One3 Compute, provider-hosted with declared residency, or One3 Box where the licence permits.
Cost and latency estimator
Estimates from the tenant's workload profile before running an eval.
Adapter publishing rules
Adapters declare base model and data provenance; adapters trained on tenant data are never publishable.
Bring-your-own provider routes
Use your own provider contract and keys for a listed route.
Kill switch and post-enablement monitoring
Disable a listing or route everywhere with fallback; monitor drift, cost and latency against the eval baseline.
Workflow
Listing review and tenant enablement
- 1
Publisher submits a model or adapter with card and licence
- 2
Artifact checks run: format, malware, hash, signature, provenance
- 3
Licence record is reviewed
- 4
Safety evals and One3 business eval suites run; critical failures reject the listing
- 5
Tenant views card, licence, benchmarks and estimate
- 6
Tenant admin accepts the licence
- 7
Tenant eval runs on golden tasks in a sandbox
- 8
Governance check runs for high-risk uses
- 9
Route is enabled with data-class limits, fallback and budget
- 10
Canary runs with rollback thresholds, then promotion and monitoring
AI
AI skills and Digital Employees
Tenant evaluation orchestration
Runs golden tasks and Actions against a candidate model in a sandbox, within budget.
Evaluation report narration
Summarises results with an explicit verdict and sample sizes.
Licence term extraction
Extracts licence terms and flags conflicts with intended use; acceptance stays human.
Safety evaluation harness
Runs injection, exfiltration, PII, jailbreak and regulated-claim suites.
Drift monitoring
Monitors enabled routes and proposes containment; disabling stays with a human.
Digital Employees that use this module
- AI Agent Governance Auditor
- AI Regulatory Watch Analyst
They act through the same governed Actions as people; risky steps wait for approval.
Frequently asked questions
Can I enable a model without testing it?
No. The design requires a tenant evaluation report on your own golden tasks before a route is enabled.
Are benchmark numbers verified?
Benchmarks declared by publishers are labelled "publisher-reported". One3 also runs its own synthetic business eval suites and your tenant evals.
Can models run on-premises?
Yes, on One3 Box where the model's licence permits on-prem use.
Can a publisher list an adapter trained on my data?
No. Adapters trained on any tenant's data are never publishable.
See Model Marketplace 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.