diff --git a/mlx/backend/cpu/simd/accelerate_fp16_simd.h b/mlx/backend/cpu/simd/accelerate_fp16_simd.h
index 1f21d2e180..950544895d 100644
--- a/mlx/backend/cpu/simd/accelerate_fp16_simd.h
+++ b/mlx/backend/cpu/simd/accelerate_fp16_simd.h
@@ -17,7 +17,7 @@ struct ScalarT<float16_t, N> {
 #endif
 
 template <>
-static constexpr int max_size<float16_t> = N;
+inline constexpr int max_size<float16_t> = N;
 
 #define SIMD_FP16_DEFAULT_UNARY(op)                    \
   template <>                                          \
diff --git a/mlx/backend/cpu/simd/accelerate_simd.h b/mlx/backend/cpu/simd/accelerate_simd.h
index a14d99103c..37b3cdbd8b 100644
--- a/mlx/backend/cpu/simd/accelerate_simd.h
+++ b/mlx/backend/cpu/simd/accelerate_simd.h
@@ -83,25 +83,25 @@ struct Simd {
 // Values chosen based on benchmarks on M3 Max
 // TODO: consider choosing these more optimally
 template <>
-static constexpr int max_size<int8_t> = 16;
+inline constexpr int max_size<int8_t> = 16;
 template <>
-static constexpr int max_size<int16_t> = 16;
+inline constexpr int max_size<int16_t> = 16;
 template <>
-static constexpr int max_size<int> = 8;
+inline constexpr int max_size<int> = 8;
 template <>
-static constexpr int max_size<int64_t> = 4;
+inline constexpr int max_size<int64_t> = 4;
 template <>
-static constexpr int max_size<uint8_t> = 16;
+inline constexpr int max_size<uint8_t> = 16;
 template <>
-static constexpr int max_size<uint16_t> = 16;
+inline constexpr int max_size<uint16_t> = 16;
 template <>
-static constexpr int max_size<uint32_t> = 8;
+inline constexpr int max_size<uint32_t> = 8;
 template <>
-static constexpr int max_size<uint64_t> = 4;
+inline constexpr int max_size<uint64_t> = 4;
 template <>
-static constexpr int max_size<float> = 8;
+inline constexpr int max_size<float> = 8;
 template <>
-static constexpr int max_size<double> = 4;
+inline constexpr int max_size<double> = 4;
 
 #define SIMD_DEFAULT_UNARY(name, op) \
   template <typename T, int N>       \