Proxmox VE releases
What changed in each release, and what the maker says will break. Newest first.
Release notes are written for people who already run the system, so the part that matters most to everyone else tends to be buried: the handful of changes that stop a machine booting or stop a virtual machine starting. Those are pulled out here, in plain words, with who they affect.
Taken from the maker's own release history and checked on 21 September 2026. Covering 8 releases and 32 recorded breaking changes. See also requirements for the current release.
Proxmox VE 9.2
Adds automatic load balancing of HA-managed VMs and containers across cluster nodes.
- Cluster Resource Scheduler can automatically rebalance HA-managed guests across nodes using live resource-usage metrics.
- Software-Defined Networking adds WireGuard and BGP fabrics, plus route maps and prefix lists for filtering BGP/EVPN routes.
- Custom CPU models can now be created, edited and removed directly from the web interface.
- The HA stack can be disarmed and re-armed cluster-wide for safer maintenance windows.
What breaks in 9.2
- Transient upgrade stall with a disarmed HA stack. If you disarm HA for cluster-wide maintenance and then upgrade while HA resources are still migrating, the upgrade can stall on the pve-ha-manager package. Affects: Clusters that disarm HA during an upgrade on a version before 5.2.4; enterprise-repository users were never affected, and simply re-arming HA lets a stuck upgrade continue.
- OVH-hosted systems can boot into memtest86+ instead of Proxmox. 9.2 adds a memtest86+ boot entry; on OVH servers managed by proxmox-boot-tool but not installed from OVH's own template, the boot process can pick memtest86+ instead of Proxmox VE. Affects: OVH customers using proxmox-boot-tool (ZFS/BTRFS root) who installed via IPMI rather than OVH's install template.
- Windows VMs with Virtualization-based Security can freeze on Intel hosts. Windows 11/2022/2025 guests using the host CPU type with VBS enabled can intermittently freeze, showing 100% CPU usage in the VM summary while frozen. Affects: Windows guests on Intel hosts using the host CPU type with VBS; fixed by upgrading qemu-server to 9.2.0 and moving affected VMs to machine version 11.0+pve2 or newer.
Proxmox VE 9.1
Lets you create LXC containers directly from OCI (Docker-style) container images.
- LXC containers can now be created from OCI images, uploaded manually or pulled from a registry.
- Virtual TPM state can be stored in qcow2 format, enabling snapshots of VMs with a TPM on file-based storage.
- Fine-grained control over nested virtualization for VM guests, useful for nested hypervisors or Windows VBS.
- SDN status reporting in the GUI is more detailed: bridges/VNets show connected guests, EVPN zones show learned IPs and MACs.
- Kernel 6.17 becomes the new stable default kernel.
What breaks in 9.1
- NVIDIA vGPU driver must be updated before upgrading. Kernel 6.17 requires at least NVIDIA vGPU/GRID driver version 19.4; upgrading to 9.1 with an older driver breaks vGPU support. Affects: Anyone running NVIDIA GRID/vGPU; must update the driver first, or pin an older kernel after upgrading and unpin once the driver is updated.
- Some Dell PowerEdge servers fail to boot kernel 6.17. Certain Dell PowerEdge models report machine-check errors or fail to boot on kernel 6.17, while kernel 6.14 still boots successfully. Affects: Specific Dell PowerEdge hardware; workaround is enabling SR-IOV Global and I/OAT DMA in firmware, or pinning kernel 6.14.
- LINSTOR/DRBD not yet compatible with kernel 6.17. The DRBD kernel module fails to build via DKMS against kernel 6.17. Affects: Users running LINSTOR/DRBD storage; workaround is manually installing and pinning kernel 6.14 until a fix is available.
- AMD Zen 5 RDSEED instruction bug on kernel 6.17.9+. Affected systems disable the RDSEED32 CPU flag at boot, which then makes QEMU guests using CPU types that expose that bit (e.g. EPYC, EPYC-v4) fail to start unless the BIOS/microcode is updated. Affects: AMD Zen 5 systems on kernel 6.17.9 or newer without an updated amd64-microcode package or BIOS; can be mitigated by using an unaffected CPU type or, as a last resort, a kernel command-line flag.
Proxmox VE 9.0
Major new release moving the whole platform onto Debian 13 Trixie, with a seamless upgrade path from 8.4.
- New major release based on Debian Trixie, with a seamless in-place upgrade from Proxmox VE 8.4.
- VM snapshots now work on thick-provisioned shared LVM storage via a new 'snapshots as volume chains' mode (technology preview).
- New HA affinity rules let you pin guests to specific nodes, or keep/split sets of guests across different nodes.
- Software-Defined Networking gains 'Fabrics', routed OpenFabric/OSPF networks between Proxmox nodes, usable for full-mesh Ceph or as a VPN underlay.
- Modernized mobile web interface, shown automatically when the GUI is accessed from a phone or tablet.
What breaks in 9.0
- Test repository renamed from pvetest to pve-test. Systems still referencing the old 'pvetest' repository name in their sources list need to update it to 'pve-test'. Affects: Anyone using the no-subscription test repository.
- Network interface names can change on upgrade. Upgrading the kernel (e.g. from 6.8 to 6.14) can change NIC names; if the previous name isn't kept as an alternative name, network configuration and firewall rules referencing it stop working until manually fixed. Affects: Hosts upgrading from Proxmox VE 8.x to 9.0, especially with less common NIC hardware; the pve-network-interface-pinning tool can pin names to custom values before upgrading.
- VirtIO virtual NIC default MTU behaviour changed. A VirtIO vNIC with no MTU explicitly set now inherits the bridge's MTU instead of defaulting to 1500, which can silently change a guest's effective network MTU. Affects: VMs with VirtIO NICs left at the previous default MTU; the pve8to9 checklist script flags affected vNICs before upgrade.
- AppArmor upgraded to version 4.1. Some packages outside the core Proxmox VE distribution (for example the CUPS printing daemon) can regress under the newer AppArmor profiles, and there is a known issue with nested containers (Docker inside an LXC container). Affects: Hosts running AppArmor-confined services, or Docker-in-LXC setups.
- VM.Monitor privilege removed. The VM.Monitor privilege is dropped in favour of Sys.Audit plus more fine-grained guest-agent privileges, so custom roles built around VM.Monitor lose the access they used to grant. Affects: Admins with custom RBAC roles referencing VM.Monitor; the pve8to9 checklist script detects affected roles.
Proxmox VE 8.4
Adds live migration for VMs using mediated devices such as NVIDIA vGPU.
- Live migration now works for VM guests using mediated devices such as NVIDIA vGPU, subject to hardware and driver support.
- A new backup-provider API lets external backup vendors write plugins that integrate with the Proxmox VE backup stack and GUI.
- VM guests can share host directories via virtiofs, without the overhead of a network filesystem.
- Linux kernel 6.14 is available as an opt-in kernel alongside the 6.8 default.
What breaks in 8.4
- PXE boot on OVMF VMs now requires a VirtIO RNG device. OVMF firmware now blocks PXE boot for guests without a random number generator, so PXE silently stops working until a VirtIO RNG device is added. Affects: Any VM using OVMF (UEFI) firmware together with PXE boot.
- iGPU passthrough broken in legacy mode. Passing an integrated GPU through to a VM fails when using legacy mode (legacy-igd=1) with the i440fx machine type. Affects: Users passing through an integrated GPU on i440fx; switching to the q35 machine type or non-legacy mode avoids the issue.
- Newly created Ceph Squid OSDs crash. OSDs created on Ceph 19.2.0 through 19.2.1-pve2 crash, particularly affecting erasure-coded pools. Affects: Ceph Squid clusters creating new OSDs; fixed in Ceph 19.2.1-pve3 or by manually changing the bluestore_elastic_shared_blobs setting.
- "Download from URL" now uses the configured proxy for HTTPS too. ISO/template/OVA downloads that previously bypassed the datacenter proxy for HTTPS connections now route through it, which can break downloads from internal HTTPS repositories the proxy cannot reach. Affects: Hosts with a datacenter proxy configured that download files via HTTPS directly from an internal-only repository.
Proxmox VE 8.3
Adds a 'Tag View' for grouping and quickly reviewing virtual guests by tag.
- New 'Tag View' groups virtual guests by their assigned tags for a quicker categorized overview.
- Software-Defined Networking is tied more tightly to the firewall, generating IP sets from VNets/guests for use in firewall rules.
- Guest import from OVF/OVA files is more streamlined, including direct upload of OVA files from the local machine.
What breaks in 8.3
- IPAM backing-path change during cluster upgrade (technology preview). IPAM state changes made on nodes that are not yet upgraded during a rolling cluster upgrade are lost, because the IPAM database files move to a new location when the first node's libpve-network-perl package updates. Affects: Clusters using the technology-preview Proxmox VE IPAM with DHCP-enabled SDN VNets, only during the upgrade window itself.
Proxmox VE 8.2
Adds a new import wizard for migrating guests directly from other hypervisors, starting with VMware ESXi.
- New import wizard connects to another hypervisor's public API (first supporting VMware ESXi) and migrates guests directly into Proxmox VE.
- Automated and unattended installation is now supported via an answer file provided on the ISO, a USB drive, or over the network.
- New 'backup fleecing' feature uses fast local storage as a buffer to reduce the I/O impact of backing up running, I/O-heavy VMs to a slow target.
- Opt-in technology-preview firewall implementation based on nftables, written in Rust.
What breaks in 8.2
- Kernel 6.8 becomes the new default. A major kernel version change can cause hardware-specific side effects. Affects: All hosts upgrading to 8.2, particularly those with less common hardware; can be avoided by pinning the proxmox-default-kernel package to version 1.0.1 before upgrading.
- Network interface names can change. NICs, especially models using the i40e driver, can be renamed (for example gaining an added port suffix) after the kernel upgrade, breaking existing network configuration. Affects: Hosts with i40e-based NICs; pinning interface names to MAC addresses beforehand avoids the problem.
- DKMS kernel modules may fail to build. Third-party kernel modules, such as the NVIDIA GRID 17.1 driver as of April 2024, fail to build against the new kernel 6.8. Affects: Hosts relying on DKMS-built drivers; workaround is pinning the proxmox-default-kernel package to the older version before upgrading.
- Split lock detection can slow down VMs. On host CPUs that support split lock detection, a VM performing misaligned memory accesses can be artificially throttled by the host kernel (kernel 5.19 or later). Affects: VMs performing misaligned memory access on affected CPUs.
- intel_iommu now defaults to on. Enabling IOMMU by default in the 6.8 kernel series can cause boot or stability problems on older hardware or systems with an outdated BIOS. Affects: Older Intel systems with buggy or out-of-date BIOS firmware; fixed by explicitly disabling intel_iommu on the kernel command line.
Proxmox VE 8.1
Adds Secure Boot support via a Proxmox-signed shim bootloader.
- Secure Boot is now supported end-to-end, including a signed shim bootloader trusted by most hardware's UEFI implementations.
- The Software-Defined Networking stack moves from experimental to fully supported, including a new automatic DHCP IP address management (IPAM) plugin.
- New flexible notification system for sending alerts through multiple channels.
What breaks in 8.1
- ZFS plus hugepages kernel bug on 6.5. Starting a VM configured with hugepages on kernel 6.5 with ZFS can hit a kernel bug that requires rebooting the host. Affects: Hosts running ZFS together with VMs configured to use hugepages.
- Some Intel Wi-Fi cards fail to initialize on kernel 6.5. Cards such as the Intel AX201 can fail to initialize under the new kernel, which can cut off remote access if Wi-Fi is the server's only network connection. Affects: Hosts relying solely on an affected Intel Wi-Fi card for network access; avoid booting into the new kernel without physical access to the machine.
- AMD fTPM hardware RNG disabled. The hardware random-number generator is disabled on all AMD systems equipped with a firmware-based TPM, affecting anything reading from /dev/hwrng. Affects: AMD systems with an fTPM device; affected systems should switch their RNG source to an alternative such as /dev/urandom.
- Dell BCM5720 NIC compatibility issue. Certain 14th-generation Dell servers have compatibility problems with the tg3 driver under the kernel 6.5.11-based build. Affects: Specific Dell server models (T140, R240, R640 and similar); workaround is pinning the kernel to 6.2 or disabling X2APIC in the BIOS.
- Default network configuration switches from ifupdown to ifupdown2. Systems installed on top of Debian, or installed before Proxmox VE 7.0, switch to the ifupdown2 network implementation; IPv6 router-advertisement acceptance (accept_ra) now defaults to off instead of on when no gateway is configured, which can silently drop IPv6 auto-configuration. Affects: Upgraded (non-fresh-ISO) installs that relied on accepting router advertisements without a gateway configured; accept_ra 2 can be added back manually to the interface configuration.
Proxmox VE 8.0
New major release based on Debian Bookworm, adding resource mappings for PCI/USB passthrough devices.
- New major release based on Debian Bookworm, with a seamless upgrade path from Proxmox VE 7.4.
- Resource mappings let you grant a user access to specific PCI/USB devices via the API and web UI without giving them full root access.
- New SDN ACL permission path lets you grant fine-grained usage permissions for specific host bridges or VNets.
- New VM CPU models based on the x86-64 psABI microarchitecture levels, with x86-64-v2-AES as the new default for VMs created via the web UI.
What breaks in 8.0
- Minimum supported browser versions raised. The web UI now uses newer JavaScript, CSS and HTML features and requires Firefox 79+, Chrome/Chromium 85+, or Safari 14+; older browsers may show a broken or non-functional interface. Affects: Anyone accessing the web UI with an older browser version.
- Storage content types must each use a separate directory. Storage activation now fails on setups where different content types (for example ISO images and backups) were mapped to the same directory, or where content directories were manually symlinked to a common location. Affects: Storages with a custom content-dirs configuration that shares directories across content types.
- Kernel 6.2 live-migration hang. Incoming live migrations of a VM with a restricted CPU type (such as qemu64) from certain Intel-CPU source hosts still on kernel 5.15 or older can hang and pin one or more vCPUs at 100% usage. Affects: Migrations from older Proxmox VE 7.4 hosts onto 8.0 hosts, for VMs using a restricted guest CPU type; fixed in a later 6.2-based kernel build (pve-kernel-6.2.16-4-pve).
- Degraded KSM performance on multi-socket NUMA hosts. Kernel Samepage Merging (memory deduplication) performs worse on 6.2-based kernels on multi-socket NUMA systems, which can noticeably increase memory usage. Affects: Multi-socket NUMA hosts relying on KSM; the recommendation at the time was to stay on Proxmox VE 7 (with its 5.15-based kernel) until a fix landed.
- QEMU 8.0 removes some deprecated features. Guest configurations using the old '-chardev tty' or '-chardev parport' aliases, or automation calling the deprecated QMP commands get_link_status, snapshot-drive or delete-drive-snapshot, stop working. Affects: Setups with custom 'args' entries in guest configs, or scripts/automation that call the removed QMP commands.