diff --git a/src/file.S b/src/file.S
index b8da07d..76ccae6 100644
--- a/src/file.S
+++ b/src/file.S
@@ -1,53 +1,76 @@
 .data
-cat_buf:        .space 4096
-endings_buf:    .space 4097
+mod_buf_0:      .space 36864
+mod_buf_1:      .space 36864
 
 .section .text
-.globl cat_file, open_fd, cat_stdin
+.globl cat_file, open_fd
 
 # Cat out the given file to stdout
-# @param[in] a0: ptr to null-terminated file str.
+# @param[in] a0: ptr to null-terminated file str. 0 = stdin
 # @param[out] a0: result code.
 cat_file:
-    addi sp, sp, -24
-    sd s0, 0(sp)
+    addi sp, sp, -32
+    sd s0, 0(sp) 
     sd ra, 8(sp)
-    sd s1, 16(sp)
+    sd s1, 16(sp) # next modbuf to WRITE TO (inverse is read) 0=modbuf_0, 1=modbuf_1
+    sd s2, 24(sp) # current modbuf byte count
 
     mv a1, a0 # since the function takes in a0, move it to a1 for syscall
+    li s1, 1
+    li s2, 0
     
-    addi sp, sp, -8
-    sd ra, 0(sp)
+    beqz a0, 1f
     call open_fd
-    ld ra, 0(sp)
-    addi sp, sp, 8
+    j 2f
+    1:
+    li s0, 0 # stdin fd
+    2:
 
     mv s0, a1 # put fd in s0
-    la t1, show_ends
-    lb s1, 0(t1) # store if we show ends into s1
 
     beqz a0, .LReadFileBlock # if we don't have an error, jump to reading
     li a0, 1
     j .LCatFileRet
 
-
     .LReadFileBlock:
     mv a0, s0
-    la a1, cat_buf
+    la a1, mod_buf_0
     li a2, 4096
     li a7, 63
     ecall
     blez a0, .LCloseFD # if we haven't read any bytes, EOF, close FD
-    beqz s1, .LNoEndingsWrite
-    mv a1, a0
-    la a0, cat_buf
-    call _stdout_add_endings
-    j .LReadFileBlock
+    mv s2, a0
+    li s1, 1
 
-    .LNoEndingsWrite:
-    mv a2, a0 # load the amount of bytes read in previous step into bytes needed to write
+    la t1, show_ends
+    lb t1, 0(t1)
+    beqz t1, .LNoEndings
+
+    mv a1, s2 # load mod buf size to a1
+
+    beqz s1, 1f
+    la a0, mod_buf_0 # load mod_buf_0 as read buf
+    la a2, mod_buf_1 # load mod_buf_1 as write buf
+    j 2f
+    1:
+    la a0, mod_buf_1 # load mod_buf_1 as read buf
+    la a2, mod_buf_0 # load mod_buf_0 as write buf
+    2:
+
+    call _add_endings
+    xori s1, s1, 1 # flip the mod buf to write to
+    mv s2, a1 # update the current byte count
+    .LNoEndings:
+
+    beqz s1, 1f 
+    la a1, mod_buf_0
+    j 2f
+    1:
+    la a1, mod_buf_1
+    2:
+
+    mv a2, s2 # load the amount of bytes read in previous step into bytes needed to write
     li a0, 1
-    la a1, cat_buf
     li a7, 64
     ecall
     j .LReadFileBlock
@@ -61,46 +84,8 @@ cat_file:
     ld s0, 0(sp)
     ld ra, 8(sp)
     ld s1, 16(sp)
-    addi sp, sp, 24
-    ret
-
-# Cat out stdin to stdout
-# @param[in] a0: Not used.
-# @param[out] a0: Not used.
-cat_stdin:
-    addi sp, sp, -16
-    sd ra, 0(sp)
-    sd s0, 8(sp)
-
-    la t1, show_ends
-    lb s0, 0(t1) # store if we show ends into s1
-
-    .LReadStdinBlock:
-    li a0, 0
-    la a1, cat_buf
-    li a2, 4096
-    li a7, 63
-    ecall
-    blez a0, .LCatStdinRet # if we haven't read any bytes, EOF, close FD
-
-    beqz s0, .LNoEndingsStdinWrite
-    mv a1, a0
-    la a0, cat_buf
-    call _stdout_add_endings
-    j .LReadStdinBlock
-
-    .LNoEndingsStdinWrite:
-    mv a2, a0 # move bytes read into bytes needed to write
-    li a0, 1
-    la a1, cat_buf
-    li a7, 64
-    ecall
-    j .LReadStdinBlock
-
-    .LCatStdinRet:
-    ld ra, 0(sp)
-    ld s0, 8(sp)
-    addi sp, sp, 16
+    ld s2, 24(sp)
+    addi sp, sp, 32
     ret
 
 # Open fd
@@ -128,8 +113,10 @@ open_fd:
 # Print to stdout while adding "$" characters.
 # @param[in] a0: ptr to null-terminated str.
 # @param[in] a1: length of str.
+# @param[in] a2: mod buf.
 # @param[out] a0: result code.
-_stdout_add_endings:
+# @param[out] a1: bytes written to mod buf.
+_add_endings:
     addi sp, sp, -72
     sd s7, 64(sp)
     sd s6, 56(sp)
@@ -147,7 +134,7 @@ _stdout_add_endings:
     li s2, 0 # set the cat_buf char counter to 0
     li s3, 0 # set the endings_buf char counter to 0
     
-    la s5, endings_buf
+    mv s5, a2
     .LStdoutEndingsProcessChunk:
     add s4, s0, s2 # Calculate the offset address
     lb t6, 0(s4)
@@ -161,14 +148,6 @@ _stdout_add_endings:
     addi s4, s4, 1
     sb t5, 0(s4) # endings_buf[s2] = \n
     addi s3, s3, 2
-
-    li a0, 1
-    la a1, endings_buf
-    mv a2, s3
-    li a7, 64
-    ecall
-    call _zero_endings_buf
-    li s3, 0 # reset endings_buf char counter
     j 2f
 
     1:
@@ -179,15 +158,9 @@ _stdout_add_endings:
     2:
     addi s2, s2, 1 # inc cat_buf pointer
     blt s2, s1, .LStdoutEndingsProcessChunk
-    beqz s3, .LSkipEndingsPrint
-    li a0, 1
-    la a1, endings_buf
-    mv a2, s3
-    li a7, 64
-    ecall
-    .LSkipEndingsPrint:
 
     li a0, 0
+    mv a1, s3
 
     ld s0, 0(sp)
     ld s1, 8(sp)
@@ -200,16 +173,3 @@ _stdout_add_endings:
     ld s7, 64(sp)
     addi sp, sp, 72
     ret
-
-_zero_endings_buf:
-    li t0, 0
-    li t1, 4097
-
-    .LZeroEndingsBufByte:
-    la t2, endings_buf
-    add t2, t2, t0
-    sb x0, 0(t2)
-    addi t0, t0, 1
-    blt t0, t1, .LZeroEndingsBufByte
-
-    ret
diff --git a/src/main.S b/src/main.S
index 3540ebb..def4c09 100644
--- a/src/main.S
+++ b/src/main.S
@@ -19,7 +19,7 @@ help:
 .equ HELP_LEN, . - help
 
 version:
-    .ascii "cat 1.1.0\n"
+    .ascii "cat 1.2.0\n"
     .ascii "Copyright (C) 2026 lexzach\n"
     .ascii "License MIT\n"
     .ascii "This is free software: you are free to change and redistribute it.\n"
@@ -61,9 +61,9 @@ _start:
     addi sp, sp, -8
     sd a0, 0(sp)
     call _is_known_flag
-    bnez a1, .LSkipArg # check if arg is flag
     ld a0, 0(sp)
     addi sp, sp, 8
+    bnez a1, .LSkipArg # check if arg is flag
     call cat_file
     addi s2, s2, 1
     bnez a0, .LReturnFileErr # error if failed
@@ -74,7 +74,8 @@ _start:
     j .LMainReturnOk
     
     .LProcessStdin:
-    call cat_stdin
+    li a0, 0
+    call cat_file
     j .LMainReturnOk
     
     .LReturnFileErr: