KServe
v0.20.0AI & MLKServe v0.20.0 is a feature-heavy release centered on LLMInferenceService enhancements (routing gates, traffic splitting, distributed tracing, KV cache offloading) plus one medium-severity Starlette security fix. A migration of InferencePool and InferenceObjective CRDs from KServe to the llm-d.ai API group requires existing manifests to be updated.
securityStarlette Content-Type vulnerability fixed
Starlette dependency vulnerability (CVE-2026-48710) fixed in v0.20.0 by bumping to >=1.0.1 and setting Content-Type headers correctly.
breakingInferencePool and InferenceObjective CRDs moved to llm-d.ai
InferencePool and InferenceObjective CRDs have migrated from the KServe API group to the llm-d.ai API group. Existing InferencePool and InferenceObjective manifests must be updated to reflect the new API group, and any tooling that references these CRDs must be reconfigured.
breakingInferenceObjective API group changed
InferenceObjective API version has changed following CRD migration. Manifests and API clients using InferenceObjective must update to the new API group.
Key changes (7)
- Security: Starlette vulnerability (CVE-2026-48710) fixed
- Breaking: InferencePool and InferenceObjective CRDs migrated to llm-d.ai API group
- LLMInferenceService routing: model-based gates, traffic splitting, canary rollout support
- LLMInferenceService tracing: distributed tracing API added
- LLMInferenceService storage: KV cache CPU offloading with multi-tier filesystem support
- New serving runtimes: vLLM and AutoGluon Server
- Other additions: confidential model serving, Managed DRA, Anthropic Messages API support, OCI+native:// ImageVolume mounting, latency predictor sidecar injection
A weekly email arrives when a release needs action. Like the security patches and breaking changes in this release.