One API across every provider
Publish a model once and get a stable, versioned endpoint. Which cloud, accelerator or serving runtime is behind it becomes a decision you can change without touching a single caller.
MAUVI · CONTROL PLANE
Run open-weight models like managed infrastructure.
Mauvi is a multi-cloud, multi-tenant control plane that turns open-weight models into managed inference endpoints — with access control, quotas, cross-provider failover, metrics-driven autoscaling and canary releases.
2 routes serving
1 degraded
p99 812 ms
What Mauvi gives you
ENDPOINTS
ACCESS
RBAC and quotas per tenant
Teams, projects and service identities with scoped roles. Set request, token and spend ceilings per tenant, and get alerts before a ceiling turns into an incident.
RESILIENCE
Cross-provider failover
Declare an ordered set of routes. When a route degrades against its health budget, traffic shifts to the next one and trickles back only after recovery is confirmed.
CAPACITY
Metrics-driven autoscaling
Scale on queue depth, time-to-first-token and accelerator utilisation rather than CPU averages, with warm floors for the paths that have a user waiting.
RELEASES
Canary and one-step rollback
Shift a defined slice of traffic to a new weight or config, hold it against your own guardrail metrics, then promote or roll back. Rollback is a single action, not a redeploy.
VISIBILITY
Metrics, traces and cost
Latency percentiles, token throughput, error taxonomy and spend, broken down by tenant, endpoint and provider. Export to your existing stack rather than replacing it.
How a release works
Promotion is a sequence, and Mauvi holds you at each stage until the previous one has earned it.
Register the candidate
Point Mauvi at the weights and the serving config. It builds the artefact once and stages it on every provider the endpoint is allowed to use.
Warm and verify
Replicas come up on the standby providers, run canary prompts and report time-to-first-token before any real traffic is offered to them.
Shift a slice
Send one percent, then five, then twenty. You choose the increments and the minimum soak time between them.
Hold against guardrails
Error rate, latency percentiles and any custom metric you publish. Breaching a guardrail pauses the rollout automatically and pages the endpoint owner.
Promote or roll back
Promotion moves the remaining traffic and retires the old version on a delay. Rollback returns every request to the previous version in one action, with the old replicas still warm.
Multi-tenant from the first line, not the enterprise tier
Tenancy bolted on late leaks: shared quota pools, audit logs that cannot attribute an action, blast radius nobody can draw. Mauvi treats the tenant as the unit that owns endpoints, budgets and keys.
That means a platform team can hand a product team a self-service endpoint without handing over the credit card or the production cluster.
- Scoped identitiesService accounts are issued per tenant per environment, with short-lived credentials and revocation that takes effect at the edge.
- Hard and soft ceilingsSoft ceilings warn and throttle. Hard ceilings stop. Both are visible to the tenant so the conversation happens before the outage.
- Attributable auditEvery policy change, rollout and revocation records who, what, when and from where — exportable in the format your auditors ask for.
payments-platform1.2M / 1.5M req
risk-research410k / 1.0M req
support-tooling940k / 1.0M req
sandbox24k / 200k req
support-tooling at 94% · soft ceiling
At a glance
| Deployment targets | AWS, Google Cloud, Azure and self-managed Kubernetes, including on-premise GPU estates. Endpoints may span any combination. |
|---|---|
| Model formats | Open-weight text, embedding and reranking models in common safetensors and GGUF distributions, served through vLLM, TGI or a custom runtime you supply. |
| Where it runs | The control plane is available as managed SaaS or self-hosted. Inference always runs in your own accounts; weights and prompts do not transit Beztly. |
| Access control | OIDC and SAML single sign-on, SCIM provisioning, role-based access down to the endpoint, and short-lived service credentials. |
| Routing policy | Ordered routes with per-route health budgets, weighted splits, regional pinning and data-residency constraints. |
| Observability | OpenTelemetry traces, Prometheus-compatible metrics, structured decision logs and per-tenant cost attribution. |
Common questions
Does our data pass through Beztly?
No. Mauvi's control plane distributes configuration and collects metrics. Prompts, completions and weights stay inside your own cloud accounts, on infrastructure you own.
What happens if the control plane is unreachable?
Data-plane routing continues with the last known good configuration. You lose the ability to make changes until connectivity returns; you do not lose the ability to serve traffic.
Can we bring our own serving runtime?
Yes. If it exposes a health endpoint and an OpenAI-compatible or gRPC interface, Mauvi can schedule, scale and route to it.
How is this different from a model gateway?
A gateway sits in front of endpoints somebody else operates. Mauvi operates them: it places replicas, scales them, releases new versions and owns the health model behind failover.
Do we have to move everything at once?
Most teams start with one endpoint on one provider, add a second provider for failover, then migrate the rest once the release workflow is familiar.
See a failover you did not have to plan for.
The fastest demo is your own workload. Bring one model and one provider and we will have a second route live in the same session.