Modelplane v0.5 is out. The fleet gateway is now an AI gateway. It authenticates callers and reads the model a request asks for, then fails over between the backends that serve it. Modelplane also collects your fleet's metrics under one set of names, whatever engine produced them. And Civo joins the clouds Modelplane can provision a cluster on.
A ModelDeployment can also scale to zero replicas now, so one you aren't
serving from costs you no GPUs. Here's what's new.
The fleet gateway is an AI gateway# #
The fleet gateway used to be an HTTP router on the control plane. It understood
nothing about the requests it forwarded. A caller reached a ModelService by
path prefix, nothing authenticated them, and the hop out to each cluster
crossed the public internet in plain HTTP.
It's now Agent Router, running on an
InferenceCluster you name rather than on the control plane, and the same one
every cluster already runs at its edge. We run a build from before the
project left the Envoy family and took that name, so you'll still see
envoy-ai-gateway-system on a cluster. It reads the model a request
names in its body and resolves the ModelService that serves it. For the
backend it picks, it rewrites the model name, the credential and the path, so a
backend sees the name it knows and a caller's key never reaches a third party.
ModelService gains priority alongside weight. Weight splits traffic
between backends at one priority; priority fails over to the next when they go
unhealthy. Every request meters a token count per caller, streams included.
You can now run more than one InferenceGateway. Each names the cluster it
runs on, so a fleet can have one per region for residency, or two in a region
for availability: Traffic from a fleet gateway to a cluster gateway is now mutually authenticated and encrypted. It used to cross the public internet as plain HTTP.
One vocabulary for a fleet's metrics# #
Modelplane doesn't own your engine. You bring the image and the command, and a
ModelDeployment runs vLLM, SGLang, or anything else that speaks the OpenAI
API, without Modelplane knowing anything about it.
That freedom is what makes a fleet hard to watch. vLLM publishes
vllm:num_requests_waiting. SGLang calls the same measurement
sglang:num_queue_reqs. DCGM, NVIDIA's
GPU exporter, reports framebuffer memory in mebibytes under a name that says
bytes, and energy in millijoules under a name that says joules.
A dashboard written against one engine is wrong on the next, and a fleet
running both has no fleet-wide number at all. Standardising on one engine would
cost you the choice that made Modelplane worth using, so Modelplane normalises
the names where it collects them instead.
Modelplane now runs an OpenTelemetry collector on every inference cluster. It
discovers every component Modelplane installs, renames each one's series into a
single modelplane_* vocabulary, and exports them wherever you say.
Two new kinds. A TelemetryDestination says where metrics go, and nothing is
collected until one exists:
A MetricMapping says what a component emits and what Modelplane calls it.
Modelplane comes with mappings for vLLM, SGLang, the gateway, the endpoint
picker and DCGM, so those need nothing from you. Write one for an engine Modelplane
has never seen and its numbers join the same surface:
fromUnit is the one to get right. A metric's name is no guide to its unit, so
the mapping says what the source measures in and Modelplane converts, histogram
buckets and all. Skip it and a series named _seconds holding milliseconds
reads a thousand times fast, with nothing downstream to catch it.
## Civo[#](#civo)
[Civo](https://www.civo.com/) joins EKS, AKS, GKE, Nebius and Vultr as a cloud
Modelplane can provision an InferenceCluster on, with the same spec you'd
write for any of them.
Two things about Civo needed handling underneath. Its GPU images carry no
NVIDIA driver, so the serving stack installs the GPU Operator to supply one,
with the toolkit and device plugin switched off so the DRA driver stays the
only thing allocating GPUs. And Civo has no server-side autoscaler, so a pool
with a maxNodeCount is scaled by the upstream cluster-autoscaler running on
the cluster itself. Civo's volumes are ReadWriteOnce, so ModelCache isn't
available there yet.
Scaling to zero# #
A ModelDeployment can now scale to zero replicas. spec.replicas used to carry
a floor of one, so kubectl scale --replicas=0 was rejected at admission. That
is awkward when scaling to zero is most of the reason to put KEDA in front of a
GPU workload in the first place. There was no way to park a deployment
either: withdrawing its endpoints while keeping the object meant tainting the
cluster hosting it.
Dropping the floor on its own would have made a parked deployment look broken. Zero desired replicas against an empty schedule reads as none of them scheduled, and on a control plane with no clusters it reads as having nowhere to run, so a deployment you had deliberately parked would sit there permanently not ready.
Zero now takes a path of its own. Nothing is composed, so the deployment's
ModelReplicas and ModelEndpoints are pruned, status.replicas reports 0 to
the scale subresource, and readiness reports true with a ScaledToZero reason,
the way a Deployment at zero replicas still reports Available. A parked
deployment reads as parked rather than as failing, which is what makes it safe
for an autoscaler to do on your behalf.
## Try it[#](#try-it)
The [getting-started guide](https://docs.modelplane.ai/getting-started/) covers
standing up a fleet, and Monitor the Fleet covers pointing telemetry at a backend you already run. Modelplane is Apache 2.0 and moving fast at
[github.com/modelplaneai/modelplane](https://github.com/modelplaneai/modelplane),
and questions are welcome in [Slack](https://slack.modelplane.ai).