Modelplane v0.7 is out, and it adds CoreWeave. A fleet can now provision CoreWeave Kubernetes Service (CKS) clusters alongside the other clouds Modelplane supports. With KubeCon in Salt Lake City next week, we spent the rest of the release fixing rough edges we found running real fleets.
CoreWeave
Set an InferenceCluster's source to CoreWeave and Modelplane provisions a
CKS cluster with its VPC, a system node pool, and the GPU node pools you list,
then installs the serving stack. Here's the platform from the
getting-started guide, one node of eight NVIDIA RTX PRO 6000 Blackwell GPUs:
apiVersion: modelplane.ai/v1alpha1
kind: InferenceClass
metadata:
name: rtxp6000-8x
spec:
description: "CoreWeave rtxp6000-8x, 8x NVIDIA RTX PRO 6000 Blackwell 96GB"
provisioning:
provider: CoreWeave
coreweave:
# CKS instance types vary by zone.
instanceType: rtxp6000-8x
accelerator:
type: nvidia-rtx-pro-6000
count: 8
devices:
- name: gpu
claim: DRA
driver: gpu.nvidia.com
deviceClassName: gpu.nvidia.com
count: 8
attributes:
architecture: { string: Blackwell }
cudaComputeCapability: { version: "12.0.0" }
capacity:
memory: { value: "97887Mi" } # RTX PRO 6000's real reported VRAM
---
apiVersion: modelplane.ai/v1alpha1
kind: InferenceCluster
metadata:
name: coreweave-us-east-04
labels:
modelplane.ai/region: us-east-04
spec:
cluster:
source: CoreWeave
coreweave:
zone: US-EAST-04A
nodePools:
- name: gpu-rtxp6000
className: rtxp6000-8x
nodeCount: 1
minNodeCount: 1
maxNodeCount: 1The ModelDeployment an ML team writes doesn't change. Any deployment whose
devices a CoreWeave pool satisfies can now run on it.
Thanks to CoreWeave for the GPU capacity we built and tested this on. We found two bugs in VPC management while testing.
Deleting a VPC right after the cluster using it could report success and leave
the VPC behind, because CoreWeave's API answered NotFound for a VPC it
wouldn't yet delete
(provider-upjet-coreweave#4).
CoreWeave's networking team says they've fixed it in their zonal API.
CoreWeave's Crossplane provider also updated a VPC with no ingress or egress settings about twice a second, for as long as the VPC existed. CoreWeave spotted the load on their side and asked us about it (provider-upjet-coreweave#6). The fix is in review at terraform-provider-coreweave#554.
Safer rollouts
An engine's Deployment started a new pod before stopping an old one. On a pool with no free GPU, the new pod waited for a GPU the old pod would never release. Modelplane now stops one copy and starts its replacement on the freed GPU, as LeaderWorkerSet and Grove already did. An engine with several copies keeps serving with one fewer, but a one-copy engine stops serving while its new pod starts.
A deployment could also report ready partway through a rollout, because its old copies still counted as ready. It now reports ready only once every copy runs the current spec.
Changing an engine from Standalone to a Leader and Worker gang left the
old Deployment running, holding the GPUs its replacement needed. Modelplane now
deletes it.
Routing during node drains
Each model replica's endpoint picker, which chooses the engine copy that serves each request, ran as one pod. The cluster's gateway fails closed without it, so draining the picker's node took the model offline until a new picker started. In one case on Nebius that took about four minutes. Modelplane now runs two pickers per replica, spread across nodes where there's room, with a disruption budget so a drain never evicts both. The cost is fewer prefix-cache hits on replicas with several engine copies.
The first ModelRoute on a new InferenceGateway could report ready while
Envoy AI Gateway was still rolling the gateway, and return HTTP 500s. The
gateway now keeps a placeholder route attached so it doesn't roll, and waits for
any rollout before it reports ready.
Cluster fixes
The rest of the fixes are mostly specific to one cloud:
- GKE taints. GKE only taints a GPU pool added to a cluster that already has a non-GPU pool, so system pods could schedule onto GPU nodes. Modelplane now taints GPU pools itself.
- EKS system pool. The LeaderWorkerSet controller and nodewright requested a full CPU core each, far more than they use. That left an NVSentinel pod unschedulable on the default system pool. They now request a quarter and half a core.
- Vultr clusters. VKE was out of stock of the system pool's plan, so every Vultr cluster failed to create. Modelplane now uses another plan.
- Long names. An
InferenceClustername of 41 characters or more pushed Helm release names past Helm's limit. - AMD GPU pools. The scheduler counted these in whole nodes, so a one-GPU replica took a whole AMD node. They now share nodes the way NVIDIA pools have since v0.6.
Inference clusters no longer run a Prometheus stack, since Modelplane's
telemetry collector scrapes engines and GPUs itself. GPU metrics now have the
deployment, replica, role, and engine labels engine metrics do, so you
can see which deployment uses which GPU. That works on EKS, AKS, Civo, Nebius,
and CoreWeave.
Docs and recipes
The new local demo runs Modelplane on two kind clusters on your laptop, with fake GPUs and models small enough for a CPU.
The getting-started guide now covers Vultr and Civo end to end, and a new recipe serves Moonshot AI's Kimi K3.1 Pro with SGLang on one 8x B300 node.
Modelplane also has a charter, which states what's in the project's scope and what isn't, and a governance model adapted from Crossplane's.
Upgrading
Modelplane doesn't keep migration paths before v1.0, so upgrading to v0.7 has
side effects. It recreates every engine workload once. Each replica stops
serving until its new pods are ready. It replaces each InferenceGateway's
Envoy Deployment, and removes the Prometheus stack from inference clusters. On
clusters created before v0.7, it also installs a second copy of some cluster
add-ons, like the cluster autoscaler, under new Helm release names.
Try it
The full release notes are on GitHub. Modelplane is Apache 2.0 and moving fast at github.com/modelplaneai/modelplane. Questions are welcome in Slack, and if you're at KubeCon next week, come find us at the Upbound booth.



