diff --git a/evm/src/cpu/kernel/asm/hash/blake/compression.asm b/evm/src/cpu/kernel/asm/hash/blake/compression.asm
index f3ac39e1..286576c6 100644
--- a/evm/src/cpu/kernel/asm/hash/blake/compression.asm
+++ b/evm/src/cpu/kernel/asm/hash/blake/compression.asm
@@ -162,7 +162,7 @@ compression_loop:
         // stack: cur_message_addr, cur_block_byte, ...
         DUP2
         // stack: cur_block_byte, cur_message_addr, cur_block_byte, ...
-        %mload_blake_word
+        %mload_blake_word_from_bytes
         // stack: m_i, cur_message_addr, cur_block_byte, ...
         DUP2
         // stack: cur_message_addr, m_i, cur_message_addr, cur_block_byte, ...
@@ -209,7 +209,6 @@ compression_loop:
         %mstore_kernel_general
         %increment
     %endrep
-
     // stack: start + 8, invert_if_last_block, t, retdest
     PUSH 0
     // stack: 0, start + 8, invert_if_last_block, t, retdest
@@ -272,36 +271,20 @@ compression_loop:
     // stack: start, retdest
     PUSH 0
     // stack: round=0, start, retdest
-    %rep 1
+    %rep 12
         // stack: round, start, retdest
         %call_blake_g_function(0, 4, 8, 12, 0, 1)
-        // %call_blake_g_function(1, 5, 9, 13, 2, 3)
-        // %call_blake_g_function(2, 6, 10, 14, 4, 5)
-        // %call_blake_g_function(3, 7, 11, 15, 6, 7)
-        // %call_blake_g_function(0, 5, 10, 15, 8, 9)
-        // %call_blake_g_function(1, 6, 11, 12, 10, 11)
-        // %call_blake_g_function(2, 7, 8, 13, 12, 13)
-        // %call_blake_g_function(3, 4, 9, 14, 14, 15)
+        %call_blake_g_function(1, 5, 9, 13, 2, 3)
+        %call_blake_g_function(2, 6, 10, 14, 4, 5)
+        %call_blake_g_function(3, 7, 11, 15, 6, 7)
+        %call_blake_g_function(0, 5, 10, 15, 8, 9)
+        %call_blake_g_function(1, 6, 11, 12, 10, 11)
+        %call_blake_g_function(2, 7, 8, 13, 12, 13)
+        %call_blake_g_function(3, 4, 9, 14, 14, 15)
         // stack: round, start, retdest
         %increment
         // stack: round + 1, start, retdest
     %endrep
-
-
-    %blake_internal_state_addr
-    %add_const(15)
-    %rep 16
-        // stack: addr, ...
-        DUP1
-        // stack: addr, addr, ...
-        %mload_kernel_general
-        // stack: val, addr, ...
-        SWAP1
-        // stack: addr, val, ...
-        %decrement
-    %endrep
-    STOP
-
     // stack: 12, start, retdest
     POP
     POP
diff --git a/evm/src/cpu/kernel/asm/hash/blake/g_functions.asm b/evm/src/cpu/kernel/asm/hash/blake/g_functions.asm
index 93b61875..243efa14 100644
--- a/evm/src/cpu/kernel/asm/hash/blake/g_functions.asm
+++ b/evm/src/cpu/kernel/asm/hash/blake/g_functions.asm
@@ -103,7 +103,7 @@
     // stack: s[y_idx], round, start
     %blake_message_addr
     ADD
-    %mload_blake_word
+    %mload_kernel_general
     // stack: m[s[y_idx]], round, start
     PUSH $x_idx
     DUP3
@@ -112,7 +112,7 @@
     // stack: s[x_idx], m[s[y_idx]], round, start
     %blake_message_addr
     ADD
-    %mload_blake_word
+    %mload_kernel_general
     // stack: m[s[x_idx]], m[s[y_idx]], round, start
     %stack (ss: 2, r, s) -> (ss, s, r, s)
     // stack: m[s[x_idx]], m[s[y_idx]], start, round, start
diff --git a/evm/src/cpu/kernel/asm/hash/blake/util.asm b/evm/src/cpu/kernel/asm/hash/blake/util.asm
index 61b731b6..916557f2 100644
--- a/evm/src/cpu/kernel/asm/hash/blake/util.asm
+++ b/evm/src/cpu/kernel/asm/hash/blake/util.asm
@@ -1,5 +1,5 @@
 // Load a 64-bit word from kernel general memory.
-%macro mload_blake_word
+%macro mload_blake_word_from_bytes
     // stack: offset
     DUP1
     %mload_kernel_general_u32_LE
@@ -7,7 +7,7 @@
     SWAP1
     // stack: offset, lo
     %add_const(4)
-    %mload_kernel_general_u32
+    %mload_kernel_general_u32_LE
     // stack: hi, lo
     %shl_const(32)
     // stack: hi << 32, lo
diff --git a/evm/src/cpu/kernel/tests/hash.rs b/evm/src/cpu/kernel/tests/hash.rs
index 0d9b8d18..de06f760 100644
--- a/evm/src/cpu/kernel/tests/hash.rs
+++ b/evm/src/cpu/kernel/tests/hash.rs
@@ -41,7 +41,9 @@ fn make_random_input() -> Vec<u8> {
 fn make_custom_input() -> Vec<u8> {
     // Hardcode a custom message
     vec![
-        1, 2, 3, 4, 5,
+        86, 124, 206, 245, 74, 57, 250, 43, 60, 30, 254, 43, 143, 144, 242, 215, 13, 103, 237, 61,
+        90, 105, 123, 250, 189, 181, 110, 192, 227, 57, 145, 46, 221, 238, 7, 181, 146, 111, 209,
+        150, 31, 157, 229, 126, 206, 105, 37, 17,
     ]
 }
 
@@ -115,15 +117,12 @@ fn test_hash_512(hash_fn_label: &str, standard_implementation: &dyn Fn(Vec<u8>)
     let random_stack = result_random.stack();
     let custom_stack = result_custom.stack();
 
-    dbg!(custom_stack);
-    // dbg!(result_custom.get_kernel_general_memory());
-
     // Extract the final output.
     let actual_random = combine_u256s(random_stack[0], random_stack[1]);
     let actual_custom = combine_u256s(custom_stack[0], custom_stack[1]);
 
     // Check that the result is correct.
-    // assert_eq!(expected_random, actual_random);
+    assert_eq!(expected_random, actual_random);
     assert_eq!(expected_custom, actual_custom);
 
     Ok(())