Skip to content

Reference

Context

Look-up facts: current versions, release history, support lifecycle, subscription pricing, system requirements, network ports, repository definitions, the storage capability matrix, CRS modes, and hard limits.

Current Versions (September 2026)

Product Version Released Base / notes
Proxmox VE 9.2 (x86-64) 2026-05-21 Debian 13.5 Trixie, kernel 7.0 (default), QEMU 11.0, LXC 7.0, ZFS 2.4; pve-manager 9.2.21 as of 2026-09-24
Proxmox VE 9.2 (arm64) 2026-08-05 First officially supported arm64 build; same feature set
Proxmox Backup Server 4.2 2026-04-29 Debian 13.4, kernel 7.0, ZFS 2.4; S3-compatible datastores now officially supported
Proxmox Datacenter Manager 1.1 2026-05-28 Debian 13.5, kernel 7.0; 1.0 shipped 2025-12-04
Ceph (integrated) Tentacle 20.2 / Squid 19.2 with PVE 9.2 Tentacle is the default for new installs since 9.2; existing clusters keep their release

PVE 8 is end-of-life

Proxmox VE 8 (Debian 12 Bookworm) reached Proxmox EOL in 2026-08. Nodes still on 8.x no longer receive updates; upgrade to 9.x (see How-to guides).

Release History (recent)

Release Date Base Notable
PVE 9.0 2025-08-05 Debian 13.0, kernel 6.14, QEMU 10.0, ZFS 2.3, Ceph Squid 19.2 Snapshots-as-volume-chain (tech preview, incl. thick LVM); HA affinity rules replace HA groups; SDN fabrics; Rust/Yew mobile UI
PVE 9.1 2025-11-19 Debian 13.2, kernel 6.17, QEMU 10.1, LXC 6.0.5, ZFS 2.3.4, Ceph Squid 19.2.3 LXC from OCI images (tech preview); vTPM state in qcow2 (snapshots with TPM); granular nested virtualization
PVE 9.2 2026-05-21 Debian 13.5, kernel 7.0, QEMU 11.0, LXC 7.0, ZFS 2.4, Ceph Tentacle 20.2 Dynamic-load CRS + automatic HA rebalancing; WireGuard and BGP SDN fabrics; custom CPU models in GUI; cluster-wide disarm-ha / arm-ha
PVE 9.2 arm64 2026-08-05 same userspace First arm64 release; NVIDIA Grace Hopper / Vera supported from launch

Support Lifecycle

PVE major Debian base First release Debian EOL Proxmox EOL
9 13 Trixie 2025-08 tba tba
8 12 Bookworm 2023-06 2026-07 2026-08
7 11 Bullseye 2021-07 2024-07 2024-07
6 10 Buster 2019-07 2022-09 2022-09

PVE follows a rolling model: each major lives as long as its Debian base, roughly 3 years plus a short overlap after the next major ships. Ceph repositories carry their own EOL: Squid 19.2 — 2026-09, Tentacle 20.2 — 2027-11 (per the PVE package-repository docs). Squid clusters should plan the move to Tentacle now.

Subscription Pricing (per occupied CPU socket / year, net EUR)

Tier Price Support
Premium EUR 1,100 Unlimited tickets, 2h response for critical issues, remote SSH support, offline updates; global 24/7 support from 2026-10-19
Standard EUR 550 10 tickets/yr, 4h response; to be added to 24/7 support in a Q4 2026 onboarding window
Basic EUR 370 3 tickets/yr, 1 business-day response (SLA unchanged)
Community EUR 120 Enterprise repo access, no ticket support (forum only)
No-subscription repo free Same packages, less tested; not recommended for production

Rules: a subscription covers every occupied socket (core count irrelevant); every cluster node must be covered and all nodes in a cluster must use the same level. All PVE features exist at every level — tiers differ only in support and repo access. Subscription keys are bound to CPU architecture: arm64 hosts need arm64 keys (sold via sales contact), and an x86 key is rejected on arm64. Proxmox North America Inc. (Kingston, Ontario) adds North American business-hours support alongside the 24/7 rollout.

Datacenter Manager licensing: PDM has no subscription of its own. A PDM instance gets enterprise-repo access and support when at least 80% of its configured remote nodes hold Basic or higher subscriptions.

System Requirements

Evaluation (min) Recommended production
CPU 64-bit x86 (Intel 64/AMD64) with VT-x/AMD-V, or arm64 (ARMv8-A) Same on x86; ARMv9-A or newer on arm64
RAM 1 GB + guest RAM 2 GB for OS/services + guests; ~1 GB per TB used storage for ZFS/Ceph
Disk Hard drive SSDs with power-loss protection; HW RAID with BBU or non-RAID for ZFS/Ceph (neither works on HW RAID)
Network 1 NIC Redundant multi-Gbit NICs; dedicated NIC for corosync
Passthrough — IOMMU: Intel VT-d / AMD-Vi (SMMU on arm64), enabled in firmware

arm64 specifics

  • Officially supported on NVIDIA Grace Hopper and NVIDIA Vera; other UEFI-based ARMv9-A hardware is best-effort. Device-tree-only boards (e.g. Raspberry Pi) are not supported; hosts must boot UEFI + ACPI.
  • VMs always boot via UEFI (AAVMF); no SeaBIOS. AMD SEV and Intel GVT-g are x86-only; no OS-level microcode package.
  • Mixed x86 + arm64 clusters are allowed, but live migration only works between same-architecture nodes; a guest only runs on a node of its own architecture.

Benchmark a node with pveperf (CPU + fsync throughput).

Network Ports

Port Protocol Service
8006 TCP Web UI + REST API (pveproxy)
22 TCP SSH (cluster actions, node joins)
5405-5412 UDP corosync cluster traffic
5900-5999 TCP VNC web console (WebSocket)
3128 TCP SPICE proxy
60000-60050 TCP Live migration (VM memory and local-disk data)
111 UDP rpcbind
25 TCP (outgoing) sendmail (notifications)

Package Repositories (9.x, deb822 format)

Repo URI Component Access
Enterprise (default after install) https://enterprise.proxmox.com/debian/pve pve-enterprise Any subscription level; most tested
No-subscription http://download.proxmox.com/debian/pve pve-no-subscription Free; testing / non-production
Test http://download.proxmox.com/debian/pve pve-test Developers / early testing only
Ceph enterprise https://enterprise.proxmox.com/debian/ceph-<release> enterprise Subscription
Ceph no-subscription http://download.proxmox.com/debian/ceph-<release> no-subscription Free

Suite: trixie for PVE 9.x; Signed-By: /usr/share/keyrings/proxmox-archive-keyring.gpg. Files live in /etc/apt/sources.list.d/*.sources (pve-enterprise.sources, proxmox.sources, ceph.sources). Disable an entry by adding Enabled: no. Legacy one-line .list files still work but apt on Trixie warns; apt modernize-sources converts them. The Debian non-free-firmware component is enabled by default on 9.x installs.

Storage Type Matrix

Type Plugin Shared Snapshots Notes
ZFS (local) zfspool no yes zvols; PVE replication (pvesr) source/target
Directory dir no yes (qcow2 / volume chain) any filesystem
BTRFS btrfs no yes tech preview
NFS / CIFS nfs/cifs yes yes (qcow2 / volume chain) file-based internal snapshots can block large VMs
LVM lvm no (yes on a shared LUN) via volume chain (9.0+, tech preview) thick provisioning
LVM-thin lvmthin no yes thin provisioning
iSCSI (kernel / libiscsi) iscsi/iscsidirect yes no raw LUNs; usually put LVM on top; multipath recommended
FC / SAS LUN (no dedicated plugin) yes via lvm on the LUN shared thick LVM is the classic SAN pattern
Ceph RBD rbd yes yes hyperconverged or external (monhost list)
CephFS cephfs yes yes file storage on Ceph
ZFS over iSCSI zfs yes yes remote ZFS appliance
Proxmox Backup pbs yes n/a (backup target) dedup, encryption, live restore

Storage Config Examples (/etc/pve/storage.cfg)

zfspool: vmdata
        pool tank/vmdata
        content rootdir,images
        sparse

rbd: ceph-external
        monhost 192.0.2.20 192.0.2.21 192.0.2.22
        pool ceph-external
        content images
        username pve-rbd

pbs: pbs-main
        datastore main
        server pbs.example.com
        username backup@pbs!pve
        fingerprint <sha256 fingerprint>
        prune-backups keep-daily=7,keep-weekly=4
        content backup

Backup Modes (vzdump)

Mode Downtime Consistency Use
stop full (VM restarted after backup starts; CT stopped for duration) highest small guests, maintenance windows
suspend short (CT: rsync then suspend + second rsync) good containers on storage without snapshots
snapshot none crash-consistent; guest-agent fsfreeze for VMs default for modern setups; fleecing optional

Fleecing: vzdump <vmid> --fleecing enabled=1,storage=<fast-local-storage> (or the backup job's Advanced tab). Live restore: qmrestore ... --live-restore 1 or the GUI checkbox (PBS sources only).

CRS Scheduling Modes (HA)

Mode Uses Since
basic (default) Number of active guests per node 7.x
static Configured CPU/memory of active guests 7.3
dynamic Average real CPU/memory usage plus configured quotas 9.2

Opt-in scheduling points: ha-rebalance-on-start (place a starting HA resource on the best node) and automatic rebalancing (9.2; per-resource opt-out via the auto-rebalance property). Always-active points: recovery after fencing and HA rule changes. Set in Datacenter -> Options -> Cluster Resource Scheduling.

Limits and Gotchas

  • pmxcfs config database: 128 MiB cap (thousands of VM configs fit).
  • Cluster size: no explicit node limit; production clusters of 50+ nodes reported (pvecm docs, 2021 statement).
  • corosync: up to 8 links (kronosnet); a single bond-backed link can mask partitions — use physically separate links.
  • Live migration: same CPU vendor recommended (cross-vendor "might work", untested), same architecture required; target node needs equal or newer PVE packages. Passed-through PCI/USB devices block live migration unless marked live-migration-capable (experimental, NVIDIA only).
  • Storage content types are per storage (content images,rootdir,...); a missing content type is a common "cannot allocate disk" cause.
  • pvecm expected N overrides quorum — break-glass only.
  • QDevice: supported for even node counts; not recommended for odd counts.
  • All nodes should run the same PVE version; mixed versions only during rolling upgrades.

Hardening Pointers (checklist form)

  • Management on a dedicated VLAN; 8006 + 22 reachable only from admin networks
  • 2FA (TOTP/WebAuthn) enforced on all admin realms; root login via PAM restricted
  • API tokens for automation; per-path RBAC instead of Administrator@pam everywhere
  • corosync on isolated NIC/VLAN (traffic is encrypted/authenticated with /etc/corosync/authkey)
  • PBS client-side encryption enabled for off-site copies; key custody documented and tested
  • Unprivileged LXC by default; nesting only where required
  • Enterprise (or Community) repo + timely updates; offline-update procedure for air-gapped sites
  • No nodes left on EOL majors (PVE 8 EOL 2026-08)
  • Regular restore tests (PBS live restore makes this cheap)

Sources