diff --git a/src/rp2350/boot_stage2/boot2_generic_03h.S b/src/rp2350/boot_stage2/boot2_generic_03h.S
index 4e14a4c7b..f7a836f11 100644
--- a/src/rp2350/boot_stage2/boot2_generic_03h.S
+++ b/src/rp2350/boot_stage2/boot2_generic_03h.S
@@ -46,6 +46,7 @@
 #error "RX delay greater than maximum"
 #endif
 
+#define CMD_RELEASE_POWERDOWN 0xAB
 #define CMD_READ 0x03
 
 // ----------------------------------------------------------------------------
@@ -60,6 +61,15 @@
 // CLKDIV and RXDELAY, and no constraints on CS max assertion, CS min
 // deassertion, or page boundary burst breaks.
 
+// Need to use direct serial mode to send SR commands. Choose a
+// conservative direct-mode divisor (5 MHz at 150 MHz clk_sys)
+// since the XIP-mode divisor may be unsafe without an RX delay.
+#define INIT_DIRECT_CSR (\
+    30 << QMI_DIRECT_CSR_CLKDIV_LSB | \
+    QMI_DIRECT_CSR_EN_BITS | \
+    QMI_DIRECT_CSR_AUTO_CS0N_BITS | \
+0)
+
 #define INIT_M0_TIMING (\
     1                      << QMI_M0_TIMING_COOLDOWN_LSB |\
     PICO_FLASH_SPI_RXDELAY << QMI_M0_TIMING_RXDELAY_LSB |\
@@ -100,6 +110,25 @@ regular_func _stage2_boot
     sw a0, QMI_M0_RCMD_OFFSET(a3)
     li a0, INIT_M0_RFMT
     sw a0, QMI_M0_RFMT_OFFSET(a3)
+
+    // Need to use direct serial mode to send commands.
+    li a1, INIT_DIRECT_CSR
+    sw a1, QMI_DIRECT_CSR_OFFSET(a3)
+    // Wait for cooldown on last XIP transfer to expire, by polling BUSY
+1:
+    lw a1, QMI_DIRECT_CSR_OFFSET(a3)
+    andi a1, a1, QMI_DIRECT_CSR_BUSY_BITS
+    bnez a1, 1b
+
+    // Send release power-down command, discard RX
+    li a0, CMD_RELEASE_POWERDOWN
+    sw a0, QMI_DIRECT_TX_OFFSET(a3)
+    jal wait_qmi_ready
+    lw a0, QMI_DIRECT_RX_OFFSET(a3)
+
+    // Disable direct mode
+    andi a1, a1, ~QMI_DIRECT_CSR_EN_BITS
+    sw a1, QMI_DIRECT_CSR_OFFSET(a3)
 #else
     push {lr}
     ldr r3, =XIP_QMI_BASE
@@ -109,11 +138,33 @@ regular_func _stage2_boot
     str r0, [r3, #QMI_M0_RCMD_OFFSET]
     ldr r0, =INIT_M0_RFMT
     str r0, [r3, #QMI_M0_RFMT_OFFSET]
+
+    // Need to use direct serial mode to send commands.
+    ldr r1, =INIT_DIRECT_CSR
+    str r1, [r3, #QMI_DIRECT_CSR_OFFSET]
+    // Wait for cooldown on last XIP transfer to expire, by polling BUSY
+1:
+    ldr r0, [r3, #QMI_DIRECT_CSR_OFFSET]
+    tst r0, #QMI_DIRECT_CSR_BUSY_BITS
+    bne 1b
+ 
+    // Send release power-down command, discard RX
+    movs r0, #CMD_RELEASE_POWERDOWN
+    str r0, [r3, #QMI_DIRECT_TX_OFFSET]
+    bl wait_qmi_ready
+    ldr r0, [r3, #QMI_DIRECT_RX_OFFSET]
+
+    // Disable direct mode
+    bics r1, #QMI_DIRECT_CSR_EN_BITS
+    str r1, [r3, #QMI_DIRECT_CSR_OFFSET]
 #endif
 
 // Pull in standard exit routine
 #include "boot2_helpers/exit_from_boot2.S"
 
+// Common functions
+#include "boot2_helpers/wait_qmi_ready.S"
+
 #ifndef __riscv
 .global literals
 literals:
diff --git a/src/rp2350/boot_stage2/boot2_w25q080.S b/src/rp2350/boot_stage2/boot2_w25q080.S
index 9d37e5f5a..71da8a1cb 100644
--- a/src/rp2350/boot_stage2/boot2_w25q080.S
+++ b/src/rp2350/boot_stage2/boot2_w25q080.S
@@ -81,6 +81,7 @@
 // A better solution is to use a volatile SR write if your device supports it.
 #define PROGRAM_STATUS_REG
 
+#define CMD_RELEASE_POWERDOWN 0xAB
 #define CMD_WRITE_ENABLE 0x06
 #define CMD_READ_STATUS 0x05
 #define CMD_READ_STATUS2 0x35
@@ -176,6 +177,12 @@ program_sregs:
     lw a1, QMI_DIRECT_CSR_OFFSET(a3)
     andi a1, a1, QMI_DIRECT_CSR_BUSY_BITS
     bnez a1, 1b
+    
+    // Send release power-down command, discard RX
+    li a0, CMD_RELEASE_POWERDOWN
+    sw a0, QMI_DIRECT_TX_OFFSET(a3)
+    jal wait_qmi_ready
+    lw a0, QMI_DIRECT_RX_OFFSET(a3)
 
     // Check whether SR needs updating
     li a0, CMD_READ_STATUS2
@@ -268,6 +275,12 @@ program_sregs:
     ldr r0, [r3, #QMI_DIRECT_CSR_OFFSET]
     tst r0, #QMI_DIRECT_CSR_BUSY_BITS
     bne 1b
+    
+    // Send release power-down command, discard RX
+    movs r0, #CMD_RELEASE_POWERDOWN
+    str r0, [r3, #QMI_DIRECT_TX_OFFSET]
+    bl wait_qmi_ready
+    ldr r0, [r3, #QMI_DIRECT_RX_OFFSET]
 
     // Check whether SR needs updating
     movs r0, #CMD_READ_STATUS2