It looks like fasmg builds up from the byte level, so it would only work for architectures that use 8-bit words. Torque builds up from the bit level, so it can assemble code for architectures like in PIC microcontrollers, using word sizes of 12 or 14 bits.
However, fasmg does allow a lot more control over the syntax of the language. The documentation shows some pretty powerful string manipulation that's used to parse real x86 assembler code, which makes sense given the purpose of it. Torque doesn't allow overriding the syntax to that degree, the macro invocation syntax is baked into the assembler.
One thing that intrigues me about fasmg is how it handles circular dependencies in expressions [0] (search for 'circular'). Currently in Torque it isn't possible to use a label reference inside a predicate, because the predicate evaluating one way could insert additional code, moving the label and causing the predicate to evaluate the other way [1]. But in fasmg it's possible to use the result of an expression as part of its own calculation.