AFAIK, a "byte" is 8 bits, but a "word" wasn't quite decided yet, and they were often divisible by 3, making octal make more sense
There were 8 address modes for each register.
So a src/dst instruction used up 12 bits for the src/dst+modes, a single register instruction used 6 bits.
You could load the registers directly at addresses 177770 through 177777 on the front panel. So if your program was 5 words long, you could load that into registers R0-R4, then load R7 with the address 177770 and execute.