diff --git a/grad.go b/internal/metal/grad.go
similarity index 99%
rename from grad.go
rename to internal/metal/grad.go
index 12ca786..55a588c 100644
--- a/grad.go
+++ b/internal/metal/grad.go
@@ -1,6 +1,6 @@
 //go:build darwin && arm64
 
-package mlx
+package metal
 
 /*
 #include "mlx/c/mlx.h"
diff --git a/io.go b/internal/metal/io.go
similarity index 99%
rename from io.go
rename to internal/metal/io.go
index 7e35773..8698dc9 100644
--- a/io.go
+++ b/internal/metal/io.go
@@ -1,6 +1,6 @@
 //go:build darwin && arm64
 
-package mlx
+package metal
 
 /*
 #include <stdlib.h>
diff --git a/lora.go b/internal/metal/lora.go
similarity index 99%
rename from lora.go
rename to internal/metal/lora.go
index ac7107e..9f9a423 100644
--- a/lora.go
+++ b/internal/metal/lora.go
@@ -1,6 +1,6 @@
 //go:build darwin && arm64
 
-package mlx
+package metal
 
 /*
 #include <stdlib.h>
diff --git a/nn.go b/internal/metal/nn.go
similarity index 99%
rename from nn.go
rename to internal/metal/nn.go
index 4de9db9..a29a75d 100644
--- a/nn.go
+++ b/internal/metal/nn.go
@@ -1,6 +1,6 @@
 //go:build darwin && arm64
 
-package mlx
+package metal
 
 // Linear is a fully-connected layer: y = x @ W.T + bias.
 // For quantized models, set Scales/Biases/GroupSize/Bits to use QuantizedMatmul.
diff --git a/optim.go b/internal/metal/optim.go
similarity index 99%
rename from optim.go
rename to internal/metal/optim.go
index f54bcf5..85c8c16 100644
--- a/optim.go
+++ b/internal/metal/optim.go
@@ -1,6 +1,6 @@
 //go:build darwin && arm64
 
-package mlx
+package metal
 
 import "math"