On August 26, 2026, the Multikernel Linux project published its first-ever public kernel tree — built on Linux 7.0 and labeled v7.0-mk2 — making it possible to run multiple fully independent Linux kernels on a single machine simultaneously, with no hypervisor and no virtual machine overhead.
The concept sounds like virtualization. It isn't. In a standard VM setup, a hypervisor creates abstraction and translation layers between the hardware and each guest operating system. Phoronix explains: Multikernel Linux eliminates those layers entirely. One "host" kernel owns the machine at boot and carves up real CPUs, physical memory ranges, and PCI devices among independent "guest" kernels, which then run directly on bare metal. There is no emulation. There are no translation layers. Each guest kernel is sovereign over its assigned hardware slice.
The implications for performance are significant. Because guest kernels access their dedicated hardware natively — not through an abstraction — workloads run at the same speed they would on a dedicated physical machine. The isolation is also strong: a crash inside one guest kernel is fully contained and won't cascade to the host or other guests.
This addresses a real problem in modern data centers. Operators often want to partition a large server into independently-managed workloads without paying the performance tax of virtualization, or the complexity cost of containers (which share a kernel and thus have weaker isolation). Hackaday noted that the project targets exactly this gap: isolation as strong as a VM, overhead nearly as low as bare metal.
The first public release is early-stage. Currently, only x86_64 is officially supported, and the project has been working on the approach for almost a year before making its code public. That said, x86_64 is the dominant architecture for servers and HPC clusters — the workloads where Multikernel Linux is most directly relevant. ARM support would open the door to cloud-native use on Graviton and Ampere processors, but that's still ahead.
The use cases are compelling: AI training runs that need bare-metal GPU access without sharing a kernel with other workloads; high-performance networking stacks that benefit from dedicated CPU cores and memory isolation; scientific computing jobs where reproducibility and isolation matter as much as raw throughput.
The Linux kernel ecosystem has seen several approaches to workload isolation over the years — containers, VMs, unikernels, cgroups. Multikernel Linux is a genuinely novel entry in that space, and the first public tree makes it possible for the broader community to evaluate the approach, file issues, and start experimenting. Worth watching closely as the project matures toward ARM support and upstream kernel submissions.