KServe

KServe is a standard model inference platform on Kubernetes, covering everything from scikit-learn models to LLMs. It is a CNCF incubating project and part of the Kubeflow ecosystem. One CNCF post dated November 2025 announces the incubation, while another tracker says it moved to CNCF as an incubator in September 2025, after starting under Kubeflow and later moving to LF AI and Data, so dates differ slightly. The GitHub releases page lists v0.21.0 as the latest, with release candidates dated late September, while the docs site still shows version 0.20, so the docs may lag the repository. One tracker lists v0.18.0 on April 29, 2026 and v0.19.0 on June 14, 2026, adding multi-node inference without Ray and OpenAI Responses API routing in 0.18, and static LoRA adapters and routing by model name in 0.19. The older modelmesh-serving repository was archived read-only on April 14, 2026.

None of the results state the licence, so I cannot confirm Apache 2.0 from this search, and you should check the LICENSE file in the kserve/kserve repository. KServe itself has no price: it is free, and your cost is the Kubernetes cluster and the GPUs or CPUs you run on, plus engineering time. No managed pricing turned up, though cloud and platform vendors may offer managed versions. Compared with a simpler framework, expect to invest in cluster operations.

KServe suits platform teams already on Kubernetes that want a standard inference protocol, autoscaling and canary rollouts. Compare BentoML, Ray and Kubeflow. Browse the MLOps & Infrastructure tools directory for more.

Model serving on Kubernetes

Standard inference.

LLM serving

Multi-node, LoRA routing.

Is KServe free?

Yes; you pay for infrastructure.

Is it a CNCF project?

Yes, incubating.

What is the latest version?

v0.21.0 on GitHub; docs show 0.20.

What licence?

Check the repository; not confirmed here.