First, a correction to our 3.0.3 coverage
When I covered 3.0.3 here on October 12, I told production users to stay on the stable 3.0.1 build, because 3.0.3 was a pre-release at the time. That advice is now stale, and it is worth saying so plainly. The GitHub listing for 3.0.3 no longer flags it as a pre-release, which means it has been promoted to the stable channel. As of today, 3.0.3 is the current stable release, and 3.0.4 is the pre-release sitting in front of it. If you stayed on 3.0.1 on my advice, running a normal wsl --update now takes you to 3.0.3, and that is the right move. The pre-release decision in front of you today is 3.0.4, not the build from last week.
What is actually in 3.0.4
Start with what is missing, because it frames everything else. The 3.0.3 changelog carried an explicit kernel update line, moving Microsoft.WSL.Kernel to 6.18.54.1. The 3.0.4 changelog has no equivalent line. The kernel you get in 3.0.3 is the kernel you keep in 3.0.4, and the work in this release lives almost entirely in the Windows-side service, the virtual machine lifecycle, networking, and the wslc container tooling.
The single largest category is memory safety and shutdown correctness in the core service, and the list is long enough to be its own story. Fixes land for potential use-after-free bugs in NatNetworking, in the stdin relay thread, in EpollWatcher, in WSLCContainer, and in the path where the WSLC virtual machine is killed while the last session reference is being released. Deadlocks are fixed between the VM exit callback and session shutdown, in WSL1 instance termination, and in WatchForExitedProcesses, which gets a combined deadlock and use-after-free fix. Hangs are fixed when a distribution is terminated while waiting on the out-of-box experience, and while waiting for the timezone thread. A potential crash during session initialization is fixed, a race in the Signal and ResizeTty process logic is fixed, and wslservice will no longer potentially kill the wslcsession process too early. Read that list as a whole and it describes a very specific kind of pain: WSL instances that refuse to die cleanly, sessions that wedge on shutdown, and a service that occasionally frees something it is still using. Anyone who has ever run wsl --shutdown and watched it do nothing knows exactly which bugs these are.
The wslc container tooling gets the most user-visible work. Containers gain volume events, so tooling watching the event stream can now see volume activity alongside container and network events, continuing the event plumbing work that 3.0.3 started with its filter consistency fixes. There is a new wslc cp alias for wslc container cp, and copying into a container now mounts bind volumes even when the container is not running, which removes an annoying start-it-just-to-copy-files step. Volume specification parsing now matches Docker's -v behavior, with a left-to-right spec parser and shared validation, another step on the Docker parity march that has defined this release line. The inspect command now handles container names that start with a slash, and a fix for the WSLC GPU hook execution and library discovery should matter to anyone running accelerated workloads in containers, because GPU containers that cannot find their libraries fail in ways that look like application bugs and are not.
Networking gets four fixes worth naming. Static DNS records no longer return incorrect AAAA results that were breaking name resolution in musl-based distributions. The NAT relay now keeps its listener alive correctly across active connections. Consomme, the networking mode that routes Linux traffic through the Windows host so it inherits VPN, proxy, and firewall behavior, now fails bind requests that fail on the Windows side instead of letting them appear to succeed. And when mirrored networking mode cannot allocate ports, it now tries smaller port ranges before giving up, which should reduce a class of mirrored-mode failures on busy hosts.
The rest is diagnostics and shutdown polish. A new /hyperv command helps triage Hyper-V issues. WSL now warns when nested virtualization capability detection fails, instead of leaving you to discover it from a cryptic downstream error. Subprocesses are reaped during shutdown to unblock systemd, and the noisy systemctl is-system-running polling errors are suppressed. A resize operation no longer treats an e2fsck return code of 1 as a failure, which is correct behavior, since that code means the filesystem check found and fixed errors rather than failing. The bundled perf integration is removed, and the default committed stack size rises to 64KB.
Who should install it, and who should wait
The action table looks a lot like the one for 3.0.3, with the version numbers shifted by one:
If you are evaluating or developing against WSL Containers, install 3.0.4 on a test machine. The volume events, the cp improvements, the Docker-matching -v parsing, and the GPU hook fix are aimed squarely at you, and the lifecycle fixes will make long container sessions less fragile.
If you are chasing a specific bug, check the fix list first. The shutdown hangs, the WSL1 termination deadlock, the musl DNS resolution failure, and the mirrored-mode port allocation problem are the kind of issues that drive people to pre-releases, and if one of them is yours, 3.0.4 is your fix today instead of at the next stable promotion.
If your WSL setup is production infrastructure, meaning developer machines your team depends on daily, stay on 3.0.3. Nothing in 3.0.4 is urgent, there is no security fix in the changelog that forces your hand, and a release this dense with concurrency fixes deserves a few days of field testing by people who opted into exactly that.
If you are still on the 2.x line and have not moved yet, 3.0.4 is not your entry point either. Move to the stable 3.0.3 build, get WSL Containers and the current platform, and let the pre-release train pass you by.
Getting it is the same drill as last time. Pre-releases do not arrive through the normal update path: run wsl --update --pre-release from an elevated prompt, or download the release package directly from the microsoft/WSL GitHub releases page, where Microsoft ships an x64 MSI, an ARM64 MSI, and the combined MSIX bundle. Check where you stand first with wsl --version, and remember that stepping back to stable is a manual move, so take the pre-release decision on machines you can afford to tinker with.
The pattern to watch
Four releases into the 3.0 line, the shape of the work is unmistakable. The 3.0.1 GA build in September brought the flagship: wslc, the container API, virtiofs, consomme. Everything since has been hardening. The 3.0.2 pre-release added vNUMA and sparse VHD experiments, 3.0.3 brought the kernel bump and stdin build contexts, and 3.0.4 is almost pure lifecycle repair. That is what a platform looks like when its team is running it hard internally and fixing what real usage shakes loose, week by week. The cadence itself, a new build roughly every week or two, says more about Microsoft's commitment to WSL Containers than any single changelog entry.
The test of the line remains the same one I named in the 3.0.3 piece: Compose. Until wsl compose up runs an existing compose.yaml unmodified, WSL Containers is a promising runtime with a missing center of gravity, and 3.0.4 does not change that status. What 3.0.4 does change is the reliability floor underneath it. A container runtime that hangs on shutdown, wedges its own service, or loses DNS in musl distributions will not displace Docker Desktop no matter how good its API is, and Microsoft is clearly spending its October fixing exactly that floor. If the next stable promotion rolls 3.0.4 up quickly, that is the signal these fixes proved out in the field. If a 3.0.5 appears within days instead, that is the other signal, and I will cover it here when it lands.