Skip to main content
Each entry lists the highlights of a Tensormesh Platform release. Use the version selector at the top of the sidebar to read the documentation for a specific release. For the exact chart, operator, and LMCache versions validated for a release, see the Compatibility matrix.
​
Security
October 18, 2026
A security release. The cache engine moves to v0.5.6, clearing a critical vulnerability and a vulnerable bundled gRPC version. Feature work is held for 1.1. If you run 1.0.0, upgrade.Upgrading from 1.0.01.0.1 ships a private coordinator image, so the chart now requires a registry credential where 1.0.0 did not. Without one the install stops at render time with an explicit error, rather than leaving pods in ImagePullBackOff. Pass --set imagePullSecret.password=<token>, or pre-create a docker-registry Secret and list it for the workload — see Install with Helm.Security
  • Cache engine v0.5.6 — clears CVE-2026-105192, an unauthenticated remote code execution reachable on the multiprocess transport, and a vulnerable bundled gRPC version.
  • Operator image v1.0.7 — rebuilt on a current Go toolchain. It carries no known vulnerability that has a published fix; the 1.0.0 operator image carried sixteen.
  • NetworkPolicy support — the operator can now create an ingress NetworkPolicy that restricts the engine server port to pods labelled for injection. Off by default, because enabling it breaks external dependencies on pod ports. Read the known issue below before turning it on.
Operations
  • tmctl pre-flight checker — no longer requires a Helm template, and its configuration now lives in its own values.yaml. Also pullable from the artifact registry as an OCI image.
Component versionsThe operator UI stays at 1.0.0 — it was not changed for this release. Platform 1.0.1 ships operator UI 1.0.0, and that pairing is the supported one. The complete artifact list, with the reference to pull each one from, is in What a release ships.Known issues
  • The NetworkPolicy has two silent preconditions. It requires a CNI that enforces NetworkPolicy, and it has no effect with hostNetwork: true. Neither raises an error, so a cluster that does not support it will accept the setting and provide no protection. Confirm your CNI enforces NetworkPolicy before relying on it as a control.
  • Isolated IPC remains off. Engine pods require hostIPC, and on OpenShift the lmcache-engine SCC.
  • CacheBlend remains Preview. Fix the recompute ratio at 0.15 for gpt-oss, and set the attention backend explicitly — automatic backend matching did not make this release.
  • On gpt-oss, set blend.checkLayer: 2. The chart defaults it to 1, which is too low for that model, and the failure is silent: blend still reuses chunks and the pods stay healthy, but a share of answers come back wrong with nothing in logs or metrics to say so. Measured on gpt-oss-120b at full reuse, 1 fails 27 of 96 retrieval trials and 2 fails 4. See CacheBlend.
​
GA
September 11, 2026
First general availability of Tensormesh Platform — self-hosted KV caching for your vLLM fleet, deployed as a Kubernetes operator.Features
  • CPU offloading (L1) — KV cache offloaded from GPU memory into host DRAM; on by default, sized by engine.spec.l1.sizeGB.
  • External storage (L2) — optional durable tier behind L1, via a filesystem adapter, Redis, or object storage, with prefetch back into L1.
  • CacheBlend (Preview) — non-prefix KV reuse for RAG and shared-block workloads.
  • KV cache encryption at rest — aesgcm encryption of the L2 tier with a master key you supply as a Kubernetes Secret.
  • P2P KV transfer (Preview) — peer-to-peer KV movement between engine pods.
  • Prefill/decode disaggregation (Preview) — split prefill and decode roles across pods.
Operations
  • Multi-tenancy — per-tenant cache isolation via cache_salt.
  • Observability — Prometheus metrics from every cache-engine pod, plus an optional OpenTelemetry Collector path for metrics and traces (Tempo).
  • tmo-cli — administrator CLI to observe and operate LMCacheEngine fleets.
  • Operator UI — a read-only console for fleet health, cache capacity, GPU telemetry, thresholds, and alerts.
Integrations
  • NVIDIA Dynamo — offload KV cache from Dynamo vLLM workers to the Tensormesh engine over cross-pod CUDA IPC.
For install and upgrade steps, see Getting Started and Install with Helm. Component versions for this release are listed in the Compatibility matrix.