Here's a small experimental release of fasm for DOS, which uses 32-bit real mode trick or DPMI if available in case of V86 mode DOS.

Attachment removed - now it's one of the official releases.

It probaly could be easily extended with the VCPI support also, so it could run in V86 even when no DPMI is available, anyone interested?

Maybe I should make it the official release for DOS? It's faster than the 16-bit FRM version currently used as a DOS stub in console release.

[little off-topic]
Privalov, there is a way to do the console version to jump to the stub code portion to avoid duplication of same code inside the exe? (sort of relocation table? or self modified code offset?)

Maybe it would be possible, but not for the current console version, because stub version has 16-bit code while the PE version uses 32-bit one, so it's cannot be reused.
And the version posted above uses 32-bit code, and runs under Win32 anyway, as Windows provides the DPMI services. It lacks only LFN support, but it also can be added, as there are some LFN services for DOS interrupt.