QEMU 10.1 is now available and comes with a handful of changes that are important to both developers and large-scale virtual machine deployers. Among the most notable features are new features in security, live migrations, architecture support, and device improvements., so it is advisable to review it calmly.
Before we get into the nitty-gritty, a note to avoid confusion: QEMU uses time-based numbering. This means that the highest release number does not, in itself, indicate the extent of the changes. Project policy allows for incompatible changes after a deprecation period with warnings in at least two previous versions., so don't assume that a major version jump is synonymous with revolution.
QEMU 10.1 Availability, Context, and Philosophy
The project announced 10.1.0 on August 26, 2025 with over 2700 commits signed by 226 authors. You can download the tarball from the official website and check the full changelog for all the details.As always, a huge thank you to those who contribute code, testing, CI, documentation, and bug reports.
In line with QEMU's philosophy, there are several mentions of removed features and deprecated options. If you depend on something older, check the 'Removed features' section and the 'Deprecated Features' chapter of the System Emulation User's Guide, where substitutions are also recommended.
Highlights at a glance
To get you started, the quick rundown of 10.1 covers security, performance, and advanced management. There is initial VFIO support for confidential guests, powerful migration improvements, and new guest agent capabilities..
- VFIO and confidentiality: Initial support for accessing/mapping sensitive guest memory when using guest_memfd, enabling passthrough on SEV-SNP/TDX.
- Live migration: multifd accelerates post-copy, pre-copy optimizations and RDMA support over IPv6.
- Guest Agent: : new 'guest-get-load' command to query load on Windows.
- Architectures: New ARM features (FEAT_SME2, ​​SME2p1, SVE2p1 and more), expanded RISC-V support and KVM integration for TDX on x86.
Architectures and machines: what exactly changes
QEMU 10.1 polishes and expands emulation and virtualization across multiple CPU families. From ARM to RISC-V and x86, through LoongArch, HPPA or Microblaze, there are concrete advances which should be broken down.
ARM
On ARM, QEMU 10.1 adds a bunch of CPU architectural features: FEAT_SME2, ​​FEAT_SME2p1, FEAT_SME_B16B16, FEAT_SME_F16F16, FEAT_SVE_B16B16 and FEAT_SVE2p1, strengthening support for vector and matrix processing.
As for machines, there are several relevant new features. The 'max78000fthr' board (Analog Devices max78000 Cortex-M4) is incorporatedA new multi-SoC, the 'AST2700FC', is coming, along with 'Catalina-BMC' (Meta) and 'GB200-BMC' (NVIDIA) machines; and firmware support is being added to the AST2700 EVB.
The popular 'virt' board is also growing: now supports CXL, PCI hotplug via ACPI and, under recent host kernels, nested virtualization with KVM by enabling 'virtualization=on'Note: The 'highbank' and 'midway' models are deprecated, and the deprecation of 'ast2700a0-evb' is announced.
RISC-V
The RISC-V ecosystem receives a barrage of details on ISA, machines, and fixes. Atomic fetch extension (Ziccif) is supported, 'Svrsw60t59b' is added, and edge cases for vector instructions are fine-tuned.
- ISA and extensions: 'tail' pseudo-instruction for queued calls, removal of CPU properties with capital 'Z', max_satp_mode from host, setting and extending PMP region count (up to 64), option to set 'vill' to reserved use of vsetvli.
- Machines: RIMT support in ACPI 'virt', customization improvements on Microchip PolarFire SoC and new Kunminghu CPU and platform; Update FADT and MADT versions in ACPI tables.
- Corrections and miscellaneous: do not allow PMP RLB to jump privileges, CSR fixes in KVM, sstc, zama16b order, profile handling, exception/masking fixes (MEPC/SEPC for IALIGN), wraparound fix on PMP 0 ranges, big endian infinite translation loop fixes, and migration failure fix when AIA is set to aplic-imsic.
Furthermore, Added missing named features, fixes to PPN translation response log, and avoided infinite delay on asynchronous transmission., which improves stability in extreme scenarios.
x86 (KVM and TCG)
On x86 there are two fine points that administrators will appreciate. Vendor-specific CPUID bits are ignored when not defined by the guest CPU vendor., avoiding confusion and artificial profiles.
On the KVM side, Added support for TDX (requires Linux 6.16 or later), and allows starting TDX or SEV-SNP machines from an IGVM file, an important step for secure provisioning flows.
In the TCG translator, TSS trap bit is supported, improving the fidelity of certain execution paths without hardware acceleration.
LoongArch
LoongArch moves forward with support from irqchip in kernel, fixes big endian with MCFG table, fixes failed startups when 'smp cpu' < 'smp maxcpus' under KVM, and polishes instruction details like VLDI. The KVM guest is also informed and a check for 'fcond' is added..
Microblaze
Those who work with Microblaze have more control: The 'petalogix_s3adsp1800' machine allows you to toggle endianness via the 'endianness' property.The big-endian variants of ml605 and xlnx-zynqmp-pmu are being deprecated, and those models are being converted to little endian in qemu-system-microblaze. Additionally, the 'qemu-system-microblazeel' binary is being deprecated and will be removed in the future.
MIPS
There is a very specific but much appreciated correction: Fixed BSOD when booting Windows NT MIPS on Magnum machine (issue #2464), useful for those who research or maintain historical software.
HPPA
Has Improved floating point exception (FPE) emulation both in user mode and system mode, reinforcing the precision of this veteran architecture.
Other architectures
In addition to the above sections, there are chapters for 68k, Alpha, AVR, Hexagon, OpenRISC, PowerPC, Renesas RX/SH, s390x, SPARC, Tricore and Xtensa. While no major headlines are listed in 10.1 for all, maintenance and tweaks continue., with special mention to s390x in recent versions for new CPU definitions and virtio-mem improvements.
Devices, I/O and subsystems
QEMU 10.1 brings important material in VFIO, networking, graphics, audio and I/O. These are changes that directly influence performance, compatibility and day-to-day operations..
VFIO
Is introduced Initial support for guest-memfd memory backends in confidential (CoCo) guests, updated IGD passthrough documentation, and fixed an L2 crash on pseries machines.
It is also enabled automatically the OpRegion for IGD passthrough, its detection is improved and QEMU can now report vfio-ap configuration changes.
Beyond that, they reach a vfio-user client, support of live update (CPR), VFIO migration with multifd on aarch64, and a property for force the PCI class code of a device. Finally, the officialization of the VFIO support on TDX and SNP VMs.
Networks and storage
There is a new backend for network devices: 'passt', designed for integration and sandboxing scenarios. Additionally, when using NBD over Unix sockets, QEMU attempts to negotiate larger buffers to improve performance dramatically.
PCI/PCIe and virtio
The ARM 'virt' board adds PCI hotplug via ACPI and acpi-index for PCI devices by enabling the global option '-global acpi-ged.acpi-pci-hotplug-with-bridge-support=on'.
In virtualized graphics, virtio-gpu allows to inject the EDID name, useful for compatibility and automation testing.
9pfs and other backends
For 9pfs, the file descriptor recovery algorithm when you are close to the system limit and several cases of use-after-unlink (operations after deleting the file) that could cause subtle errors.
Audio
The audio subsystem adjusts defaults and compatibility: In ALSA 'try-poll' is now false by default and endianness converters are added for floating point samples.
Graphical interface and agents
There is a lot to discuss in GUI: spice/dbus supports multiplane dmabuf, and gl=on can also be enabled with remote/non-local clients.
The vdagent agent adds migration support, GTK improves scale handling and adds 'keep-aspect-ratio' and 'scale' options.
In VNC they are corrected update coding issues when client and server differ in endianness, both in non-tight encodings and in 8/16 bpp; it also solves the problem screen paused after migration.
I/O and debugging
At the I/O layer, Many TCP backends now allow you to configure TCP keepalive parameters, very useful in environments with aggressive firewalls or NAT.
The GDBStub adds support for qGDBServerVersion used by non-GDB clients, improving interoperability.
TCG plugins get pampering: 'ips' plugin supports Configurable scaling and instructions per quantum; plus there are new features for write logs, write in virtual addresses (with caution) and read/write hardware addresses. A test plugin 'patcher' is incorporated, and sparc64 and alpha expose logging details to plugins.
Live migration and block tasks
10.1 tightens the screws on VM mobility. RDMA migration now supports IPv6 and postcopy is optimized for sequential accesses (typical cases such as streaming video).
Great news: Multifd and postcopy can now be enabled at the same time. (multifd is now enabled during precopy). Also fixed VFIO migration with multifd when tcp zerocopy is enabled and reduced downtime window in precopy avoiding unnecessary LOG_CLEARs.
In QMP, blockdev-mirror debuts the boolean flag 'target-is-zero' to optimize mirrors to destinations that already read as zero; furthermore, even without that flag, less work is done on source blocks that read as zero. blockdev-backup Adds the 'on-cbw-error' option to decide how to behave in the event of I/O errors during copy-before-write.
In block work, block-commit not active It is optimized to maintain sparseness; and mirror/backups inherit previous improvements for zero blocks and new options.
Guest Agent
The QEMU Guest Agent grows on Windows: 'guest-get-load' is implemented to query the guest system load.
Also in Windows a is added Retry when deregistering the VSS provider and propagates the installation error from the VSS provider to the MSI installer, making diagnosis easier.
User mode and TCG
In user-mode emulation there is minor documentation updates Clarifying limitations; areas of signals, runtime, and binfmt_misc are covered, as well as alpha, arm/aarch64, HPPA, LongArch, PowerPC, s390, and x86 targets.
Work continues at TCG on LoongArch and RISC-V and the support of Record/Replay, although with no changes to the headline beyond the aforementioned TSS cheat bit.
Compilation, dependencies and platforms
In the build process, the documentation is updated with b4 examples and all edk2 submodules now go included in the firmware tarball.
Regarding Rust, the minimum supported version increases to 1.77. On Debian, bookworm is available in the rustc-web package except for mips64el (for that purpose, Debian trixie or newer is required). On Ubuntu 22.04 and 24.04, the rust-1.77 package is used by setting the RUSTC/RUSTDOC variables to /usr/bin/rustc-1.77 and /usr/bin/rustdoc-1.77.
Rust support continues experimental and unstable, not recommended for anything other than development. With Rust enabled, QEMU requires Meson 1.8.1A script ('scripts/get-wraps-from-cargo-registry.py') is included to populate 'subprojects/' from a local registry (e.g., /usr/share/cargo/registry).
On host platforms, Debian Bullseye is no longer supported, and Ninja 1.9 becomes a requirement (it was already present on all systems supported with QEMU 10.0). It also arrives Experimental support for compiling to WASM with Emscripten.
There are no specific headlines about Windows in this installment, beyond what was already mentioned by the guest agent. The 10.1 roadmap is on the Planning/10.1 page if you want to follow the details.
If you've been waiting to upgrade, the combination of support for confidential guests (TDX/SEV-SNP with IGVM and guest_memfd), faster migrations (multifd + postcopy, RDMA IPv6), and improvements across architectures and subsystems make QEMU 10.1 a very well-rounded release. With more practical performance, better management tools, and a technological base that continues to modernize, the jump is worth it..
