Windows/386 was the first version to do that, I believe. It's worth noting that it and its descendants were based around this hypervisor architecture but the fact seems to be little-known, perhaps because the virtualisation was so "transparent" (and allowing pass-through hardware access by default). Only later did things like Xen, KVM, QubesOS, etc. "rediscover" this architecture --- but with a "deny by default" model for security instead.

IBM invented hypervisors in the 1960s, and IBM mainframes have seen continual production use of hypervisors ever since then. IBM's mainframe hypervisors had a massive impact on the computer science research community – basically all academic hypervisor research can be historically traced to that initial work inside IBM – and it was that research which the initial developers of systems such as Xen and KVM drew upon.

By contrast, Windows/386 and its successors (Win3.x 386 Enhanced Mode, and Win95/98/Me's VMM kernel) were essentially a historical dead-end. They weren't doing anything which hadn't already been done on other platforms a decade or more earlier. Systems on other platforms were doing far more advanced virtualisation at the same time. They didn't really contribute any significant ideas to future systems, or to academic research into hypervisor technology.

If Xen/KVM/QubesOS/etc started by "rediscovering" anything, it was by "rediscovering" IBM VM, not by "rediscovering" Windows/386.

Similarly, one of the cofounders of VMware, Mendel Rosenblum, is a CS professor at Stanford, and VMware’s first products were built on his academic research into hypervisors.

Academic research has had far more practical significance in this space than you give it credit for. Meanwhile, Virtual 8086 mode’s limitations meant it had limited value except as a backward compatibility environment for running legacy DOS software on newer operating systems (Windows, OS/2, Linux, etc), and as the years passed that use case became less and less relevant-V86 mode is essentially a technological dead-end. The academic research which went into VMware and Xen left a major mark on the whole industry.

And if you read academic papers on hypervisors, you’ll see IBM’s mainframe virtualisation technology cited as prior work, again and again and again.

https://en.wikipedia.org/wiki/CEMM was earlier

I had a 486 with 8Mb of RAM and a 40Mb disk, DBLSPACEd. It usually didn't had enough low memory (because DBLSPACE was a hog) for some games, so... I started Windows and played these DOS games from there.

I don't remember which one game occasionally would drop out with OOM in the DOS (it had 8Mb in the minimal requirements) but run just fine when started from Windows. Virtual memory with swap for the rescue!

I even used Calendar (if my memory serves me well) as an alarm clock to remind myself about the time.

"The Win32 API is implemented by three modules, each consisting of a 16-bit and a 32-bit component:"

Fully?