Reference

Facts

Every FlashVMM figure and count in one table.

The FlashVMM facts table is the single source of every figure and count on flashvmm.com: performance by platform, the size of every interface, and the platforms FlashVMM runs on.

FlashVMM 0.1.82

Performance

FactValueWhere
Cold boot, kernel to running18 msmacOS
Cold boot, kernel to running42 msLinux x86 and Linux x86 cloud bare metal
Cold boot, kernel to running63 msLinux ARM cloud bare metal
Create to first command output60 msmacOS
Claim a warm VM from a pool1.2 msLinux x86 cloud bare metal
Claim a warm VM from a pool2.0 msmacOS and Linux x86
Restore a snapshot to running7.8 msLinux ARM cloud bare metal
Restore a snapshot to running17 msLinux x86 cloud bare metal
Fork a running VM, child running36 msmacOS
Fork a running VM, child running38 msWindows x86
Rewind a VM to a checkpoint32 msmacOS
Wake a sleeping VM to running8.4 msLinux x86 cloud bare metal
Host memory per idle VM (256 MB given)27 MBLinux ARM
50 VMs started at once, all running125 msLinux x86 cloud bare metal
Network transfer with per-VM rules on vs off1,751 vs 1,679 Mbit/sLinux x86 cloud bare metal

Interfaces

FactValueWhere
MCP tools in flash-mcp21every platform
flash CLI commands46every platform
Per-VM REST API (flashvm)34 resources, 59 operationsevery platform
Host API (flashhostd)15 resources, 31 operationsevery platform
Fleet controller API10 resources, 17 operationsevery platform
Warm-pool API (flashpoold)5 resources, 10 operationsevery platform
GPU-pool API (flashgpud)6 resources, 11 operationsevery platform
NIC-pool API (flashnetd)5 resources, 7 operationsLinux

Platforms

FactValueWhere
LinuxKVMx86_64 and arm64
macOSHypervisor.framework and Virtualization.frameworkApple silicon
WindowsWindows Hypervisor Platformx64 and arm64
GuestLatest Linux kernel, slim busybox and musl root filesystemevery platform

AI agent sandbox

FactValueWhere
AI agent start in its sandbox (Claude Code, Enter to first output)0.33 smacOS
AI agent start in its sandbox (Claude Code, Enter to first output)0.41 sLinux x86
AI agent start in its sandbox (Claude Code, Enter to first output)0.50 sLinux ARM
AI agent start in its sandbox (Claude Code, Enter to first output)0.51 sWindows x86
AI agent start in its sandbox (Claude Code, Enter to first output)0.43 sWindows ARM
AI agent start, no warm sandbox waiting0.41 smacOS
AI agent start, no warm sandbox waiting0.42 sLinux x86
AI agent start, no warm sandbox waiting0.59 sLinux ARM
AI agent start, no warm sandbox waiting0.84 sWindows x86
AI agent start, no warm sandbox waiting0.48 sWindows ARM
One shell command in the sandbox (on the host)20 ms (8 ms)macOS
One shell command in the sandbox (on the host)14 ms (6 ms)Linux x86
One shell command in the sandbox (on the host)22 ms (11 ms)Linux ARM
One shell command in the sandbox (on the host)83 ms (43 ms)Windows x86
One shell command in the sandbox (on the host)73 ms (23 ms)Windows ARM
First use: Claude Code installed into a new sandbox24 sWindows x86 and Windows ARM

Related