On our tested Beelink EQi12, WSL2 took exactly what Microsoft's documentation says it takes: half the RAM (8 GiB assigned, 7.6 GiB reported inside), 25% of that as swap (2 GiB), a 1 TB virtual disk ceiling on the C: drive and a NAT network at 172.26.180.20/20. Every one of those measured values matched a documented default — which turns a five-command baseline check into proof of where your home server stack actually runs.
Our Docker benchmark, stability run and backup drill on the EQi12 all ran through Windows 11, WSL2 and Docker Desktop. That fact appears in the methodology of every one of those articles, but we never published the platform itself. This page is the missing layer: a point-in-time baseline of the WSL2 VM on the same machine, captured on July 11, 2026, read against Microsoft's documented defaults. The source file is EQi12_WSL_Ubuntu_Baseline.txt in the public measurement repository, released under CC BY 4.0.
Why a platform baseline matters on a mini PC
A mini PC home server has no headroom to waste. A 16 GB machine that silently hands 8 GiB to a utility VM has half its memory committed before the first container starts. A 512 GB NVMe that hosts a dynamically expanding virtual disk can fill up from inside a layer the Windows file explorer does not explain well. And a service bound inside a NAT-ed virtual network is invisible on the LAN no matter how many firewall rules you write. None of these are bugs — they are documented defaults — but every one of them changes how an unattended server behaves after a reboot or a power cut.
We hit the practical consequences elsewhere on this site: the docker stats monitoring test recorded containers seeing a 7.638 GiB memory limit, the containers-not-starting fix treats WSL2 as its own failure layer, and the Docker auto-start guide sequences Windows, login, WSL2, Docker and Compose as separate recovery stages. This baseline explains the numbers behind those guides.
The baseline snapshot, verbatim
The following values come from one capture inside the WSL2 Ubuntu environment (a few identifying details removed). Conditions: Beelink EQi12, Core i3-1215U (4 cores / 8 threads), 16 GB DDR4-3200, 512 GB WD PC SN540, Windows 11, Docker Desktop with the WSL2 backend, July 11, 2026.
| Property | Measured value |
|---|---|
| Kernel | 6.18.33.2-microsoft-standard-WSL2 (SMP PREEMPT_DYNAMIC, built Jun 18 2026) |
| Distribution | Ubuntu 26.04 LTS (Resolute Raccoon), codename resolute |
| CPU visible to WSL2 | 8 logical processors (0–7), same count as Windows; VT-x exposed, hypervisor vendor: Microsoft |
| Memory (free -h) | total 7.6 GiB · used 931 MiB · free 6.4 GiB · available 6.7 GiB |
| Swap | 2.0 GiB, 0 B used at capture |
| Root filesystem | /dev/sdf, ext4, 1007 G size, 2.6 G used (1%) |
| Windows drive | /mnt/c via 9p, 476 G with 146 G used (31%) |
| Network | eth0 UP at 172.26.180.20/20 (NAT); loopback 127.0.0.1/8 plus 10.255.255.254/32 |
| GPU bridge | /dev/dxg present (crw-rw-rw-) |
| Docker socket | present |
One snapshot proves nothing by itself. The reason this table is worth publishing is the next one.
Measured values vs documented defaults: five for five
Microsoft's advanced settings documentation for WSL (retrieved October 26, 2026) defines what an unconfigured WSL2 VM should look like. We put the documented defaults next to the measured baseline — no .wslconfig file existed on this machine, so every row is a default in action.
| Setting | Documented default | Measured on the EQi12 | Match |
|---|---|---|---|
| memory | 50% of total memory on Windows | 8 GiB assigned (16 GB host); free -h reports 7.6 GiB total | Yes, minus kernel reservations |
| swap | 25% of memory size, rounded to nearest GB | 2.0 GiB | Yes (25% of 8 GiB) |
| processors | Same number of logical processors as Windows | 8 visible (host has 8 threads) | Yes |
| defaultVhdSize | 1 TB (1,099,511,627,776 bytes) | 1007 G reported by df on /dev/sdf | Yes, ext4 overhead aside |
| networkingMode | NAT | eth0 at 172.26.180.20/20, private subnet | Yes |
Two defaults deserve extra attention because they are invisible until they cost you storage or reachability: sparseVhd (default false) and localhostForwarding (default true). They are covered in the next two sections.
Memory: the 50% ceiling your containers inherit
On a 16 GB machine, the documented 50% default assigns 8 GiB to the WSL2 VM. Inside, Ubuntu reported 7.6 GiB total — the gap is kernel and hypervisor reservations, not missing memory. What matters for a home server is that this ceiling propagates: our docker stats comparison measured the same machine's containers reporting a 7.638 GiB limit, which is this boundary showing up from the container side.
At the moment of capture, the VM held 931 MiB used with 6.4 GiB free — an idle-ish Docker Desktop stack of nginx, PostgreSQL and Redis plus the platform itself. The assignment is a ceiling rather than a constant cost, and Microsoft ships an experimental autoMemoryReclaim setting (default dropCache) that reclaims cached memory inside the VM. Still, capacity planning on a 16 GB Windows mini PC should assume roughly half the RAM is fenced off before Windows workloads and containers fight over what remains. That arithmetic is one reason our Windows+Docker vs Proxmox comparison treats memory governance as a structural difference, not a preference.
Disk: a 1 TB virtual disk that only grows
The distribution filesystem lived at /dev/sdf reporting 1007 G — the documented 1 TB default ceiling, minus ext4 formatting overhead. Only 2.6 G was used inside, while the Windows C: drive carried 146 G of 476 G. Two planning consequences follow:
- The VHDX lives on C:. Everything inside the Linux filesystem — container images, volumes, logs — consumes the same physical NVMe partition as Windows itself. The ceiling (1 TB) is larger than the underlying 512 GB SSD, so the real limit is the host drive, not the virtual disk.
- sparseVhd defaults to false. The VHDX grows as data is written and does not automatically shrink when data is deleted. Compacting requires an explicit maintenance step. For an unattended server that writes logs and media metadata, disk-free checks should watch the Windows drive, not just df inside WSL.
Networking: NAT by default, LAN by exception
The baseline's eth0 address — 172.26.180.20/20 — is a private subnet managed by Windows, separate from the physical Intel gigabit ports on the LAN. That is the documented NAT default, and it has three practical effects for a home server:
| Path | Behavior under the NAT default |
|---|---|
| Host → service | localhostForwarding defaults to true: ports bound in WSL2 are reachable from Windows via localhost |
| LAN device → service | Not directly reachable; traffic must be published through the Windows host (Docker port publishing or a relay), then Windows firewall rules apply |
| Service → LAN | Outbound works through NAT; the WSL side has no LAN IP, so MAC-binding features (Wake-on-LAN, DHCP reservations) do not apply to it |
A mirrored networking mode exists in .wslconfig and gives the WSL side the host's own addresses, but per Microsoft's documentation it requires Windows 11 22H2 or higher; bridged mode has been deprecated since WSL 2.4.5. Neither was enabled on this machine, so every measured service in our benchmark and stability articles was reached through the published-port path. The loopback entry at 10.255.255.254/32 is part of WSL's DNS tunneling plumbing rather than a second network interface.
GPU and devices: the bridge is present, the verdict is not
The device list showed /dev/dxg — the GPU paravirtualization bridge — with open permissions, consistent with the documented gpuSupport default of true. Docker Desktop's socket was present as well, which is what let the benchmark containers run from inside this VM.
We deliberately stop at the bridge. Our measured Quick Sync numbers — two concurrent 4K60-to-1080p60 transcodes at real time, four below it — come from native Jellyfin on the same machine, and the codec matrix documents the decode and encode boundaries. Whether QSV performs identically for containers inside WSL2 is a different experiment we have not run, so this page claims only: the hardware path exists at the platform layer.
Reproduce this baseline on your own mini PC
Five commands inside any WSL2 distribution, run in under a minute, produce the same evidence categories. Compare your output against the documented defaults above — any mismatch means a .wslconfig or group policy is shaping your platform.
uname -a # kernel: expect ...-microsoft-standard-WSL2
free -h # memory: ~50% of host RAM; swap: ~25% of the VM memory
df -h / # virtual disk: ~1 TB ceiling on the default VHD
ip addr show eth0 # 172.x.x.x/20 = NAT (default); a LAN IP suggests mirrored mode
ls -l /dev/dxg # GPU paravirtualization bridge (gpuSupport default true)Fact block: source and conditions
| Item | Value |
|---|---|
| Source file | wsl/EQi12_WSL_Ubuntu_Baseline.txt, eqi12-measurement-data repo (CC BY 4.0) |
| Machine | Beelink EQi12, Core i3-1215U, 16 GB DDR4-3200, 512 GB WD SN540 |
| Capture date | July 11, 2026, single point-in-time snapshot |
| Stack | Windows 11, WSL2, Docker Desktop (nginx + PostgreSQL + Redis active) |
| Documentation cross-check | Microsoft Learn WSL advanced settings, retrieved October 26, 2026 |
Boundaries of this measurement
- Single machine, single snapshot. This is one capture on one EQi12 on one morning in July 2026. It documents the platform state that our other measurements ran on top of; it is not a survey of WSL2 behavior across machines or versions.
- No performance claims. No benchmarks were run inside the WSL2 environment for this article. Throughput numbers on this site come from the dedicated network, storage and Docker benchmark pages.
- Untested features stay untested. Mirrored networking, sparse VHDX, autoMemoryReclaim tuning and QSV-in-container were not exercised on this machine; each is described here only as a documented capability.
- Version drift. The kernel string (6.18.33.2, built June 2026) and Ubuntu 26.04 identify this snapshot precisely. Newer WSL releases may change defaults; re-run the five commands to re-anchor.
What this baseline changes about running a home server in WSL2
For light stacks — a handful of containers like our nginx, PostgreSQL, Redis and Jellyfin combination — the platform is workable: the memory ceiling is predictable, the disk is bounded by the host drive, and published ports reach the LAN reliably enough that a 12.57-hour stability run completed without a sampled failure. How the ceiling turns into a real RAM budget is worked out in the 16 GB RAM breakdown, which puts container usage next to host usage. The costs are structural: half the RAM fenced off by default, a VHDX that only grows, one more reboot-recovery layer, and network identity that stops at NAT. If your ambitions trend toward VMs, ZFS or direct hardware control, that is the point where the Proxmox comparison or the native Ubuntu install guide becomes the more relevant read. If you stay on Windows, the Windows home server build guide shows where WSL2 fits into the full build, and the auto-start guide makes the platform survive reboots unattended.