diff --git a/external/ggml/src/ggml-metal/ggml-metal-device.m b/external/ggml/src/ggml-metal/ggml-metal-device.m
index dca96bc5c..b4ac18298 100644
--- a/external/ggml/src/ggml-metal/ggml-metal-device.m
+++ b/external/ggml/src/ggml-metal/ggml-metal-device.m
@@ -1279,14 +1279,14 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                         switch (op->type) {
                             case GGML_TYPE_F32:
                             case GGML_TYPE_F16:
-                                return true;
+                            case GGML_TYPE_BF16:                                return true;
                             default:
                                 return false;
                         }
                     case GGML_TYPE_BF16:
                         switch (op->type) {
                             case GGML_TYPE_F32:
-                            case GGML_TYPE_BF16:
+                            case GGML_TYPE_F16:                            case GGML_TYPE_BF16:
                                 return true;
                             default:
                                 return false;
diff --git a/external/ggml/src/ggml-metal/ggml-metal.metal b/external/ggml/src/ggml-metal/ggml-metal.metal
index b71b83c68..a2f5cf8ec 100644
--- a/external/ggml/src/ggml-metal/ggml-metal.metal
+++ b/external/ggml/src/ggml-metal/ggml-metal.metal
@@ -7932,6 +7932,8 @@ template [[host_name("kernel_cpy_contig_f16_f16")]] kernel kernel_cpy_contig_t k
 #if defined(GGML_METAL_HAS_BF16)
 template [[host_name("kernel_cpy_contig_bf16_f32")]]  kernel kernel_cpy_contig_t kernel_cpy_contig_t_t<bfloat, float>;
 template [[host_name("kernel_cpy_contig_bf16_bf16")]] kernel kernel_cpy_contig_t kernel_cpy_contig_t_t<bfloat, bfloat>;
+template [[host_name("kernel_cpy_contig_f16_bf16")]]  kernel kernel_cpy_contig_t kernel_cpy_contig_t_t<half,   bfloat>;
+template [[host_name("kernel_cpy_contig_bf16_f16")]] kernel kernel_cpy_contig_t kernel_cpy_contig_t_t<bfloat, half>;
 #endif
 
 template<typename T>
@@ -8064,6 +8066,8 @@ template [[host_name("kernel_cpy_f16_f16")]]   kernel kernel_cpy_t kernel_cpy_t_
 #if defined(GGML_METAL_HAS_BF16)
 template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  float>;
 template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  bfloat>;
+template [[host_name("kernel_cpy_f16_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<half,    bfloat>;
+template [[host_name("kernel_cpy_bf16_f16")]]  kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  half>;
 #endif
 
 template<short QK,
diff --git a/src/models/breeze_tts/generator.cpp b/src/models/breeze_tts/generator.cpp
index 4304940a5..4699a1920 100644
--- a/src/models/breeze_tts/generator.cpp
+++ b/src/models/breeze_tts/generator.cpp
@@ -55,14 +55,13 @@ struct GgmlContextDeleter {
 modules::QwenDecoderActivationCastPolicy breeze_bf16_activation_policy(core::BackendType backend_type) {
     modules::QwenDecoderActivationCastPolicy policy;
     if (backend_type != core::BackendType::Cuda && backend_type != core::BackendType::Hip &&
-        backend_type != core::BackendType::Vulkan) {
+        backend_type != core::BackendType::Vulkan && backend_type != core::BackendType::Metal) {
         return policy;
     }
     policy.enabled = true;
     policy.type = GGML_TYPE_BF16;
     // CUDA/HIP/Vulkan implement the fused round-to-bf16 unary op.
-    policy.fused_round = backend_type == core::BackendType::Cuda || backend_type == core::BackendType::Hip ||
-        backend_type == core::BackendType::Vulkan;
+    policy.fused_round = backend_type != core::BackendType::Metal;
     policy.after_input_norm = true;
     policy.after_qkv_projection = true;
     policy.after_qk_norm = true;
@@ -114,12 +113,13 @@ modules::QwenCausalDecodeRuntimeConfig backbone_config(
     out.decoder.stack.runtime.static_cache.update_mode = modules::QwenDecoderStaticCacheUpdateMode::DirectSetRows;
     out.decoder.stack.runtime.static_cache.set_rows_mode = modules::QwenDecoderStaticCacheSetRowsMode::BackendViewOptimized;
     if (backend_type == core::BackendType::Cuda || backend_type == core::BackendType::Hip ||
-        backend_type == core::BackendType::Vulkan) {
+        backend_type == core::BackendType::Vulkan || backend_type == core::BackendType::Metal) {
         // BF16 KV cache matches the reference implementation, but flash
         // attention only accelerates bf16 cache with native bf16 MMA
         // (sm_80+); on older parts it is ~3x slower, so only HIP uses it.
         out.decoder.static_cache_type =
-            backend_type == core::BackendType::Hip ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+            (backend_type == core::BackendType::Hip || backend_type == core::BackendType::Metal)
+                ? GGML_TYPE_BF16 : GGML_TYPE_F16;
         out.decoder.stack.activation_cast = breeze_bf16_activation_policy(backend_type);
     }
     out.decoder.logits_size = config.lm_head_size;
@@ -168,11 +168,12 @@ modules::QwenCausalDecodeRuntimeConfig depth_config(
     out.decoder.stack.runtime.static_cache.update_mode = modules::QwenDecoderStaticCacheUpdateMode::DirectSetRows;
     out.decoder.stack.runtime.static_cache.set_rows_mode = modules::QwenDecoderStaticCacheSetRowsMode::BackendViewOptimized;
     if (backend_type == core::BackendType::Cuda || backend_type == core::BackendType::Hip ||
-        backend_type == core::BackendType::Vulkan) {
+        backend_type == core::BackendType::Vulkan || backend_type == core::BackendType::Metal) {
         // See backbone_config: only HIP uses a bf16 KV cache; CUDA and Vulkan
         // keep F16.
         out.decoder.static_cache_type =
-            backend_type == core::BackendType::Hip ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+            (backend_type == core::BackendType::Hip || backend_type == core::BackendType::Metal)
+                ? GGML_TYPE_BF16 : GGML_TYPE_F16;
         out.decoder.stack.activation_cast = breeze_bf16_activation_policy(backend_type);
     }
     out.decoder.logits_mode = modules::QwenCausalDecoderLogitsMode::LastStep;