Add nvidia_parakeet-v3

nvidia_parakeet-v3/AudioEncoder.mlmodelc/LICENSE_NOTICE.txt

@@ -0,0 +1,7 @@
+Argmax proprietary and confidential. Under NDA.
+
+Copyright 2024 Argmax, Inc. All rights reserved.
+
+Unauthorized access, copying, use, distribution, and or commercialization of this file, via any medium or means is strictly prohibited.
+
+Please contact Argmax for licensing information at [email protected].

nvidia_parakeet-v3/AudioEncoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d7fb2cdf39ba37f092cbf481ff457b8ff84746012e4402542cd712d69a1326a
+size 243

nvidia_parakeet-v3/AudioEncoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c3a319786d1956fe24eba520a941c7e5ab8486f14558dd895d6bd9c95f45bb1
+size 493

nvidia_parakeet-v3/AudioEncoder.mlmodelc/metadata.json

@@ -0,0 +1,96 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1024 × 1 × 188)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 1, 188]",
+        "name" : "encoder_output_embeds",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 640 × 1 × 188)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1, 188]",
+        "name" : "joint_projected_encoder_output_embeds",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios16.silu" : 72,
+      "Ios17.mul" : 73,
+      "Split" : 24,
+      "Ios17.transpose" : 1,
+      "Ios17.sub" : 1,
+      "Ios17.matmul" : 72,
+      "Ios17.conv" : 295,
+      "Ios16.sigmoid" : 24,
+      "Ios17.add" : 168,
+      "Ios17.sliceByIndex" : 48,
+      "Ios17.batchNorm" : 120,
+      "Ios16.relu" : 3,
+      "Ios16.softmax" : 24,
+      "Ios17.reshape" : 193,
+      "Ios17.layerNorm" : 120,
+      "Pad" : 24
+    },
+    "computePrecision" : "Mixed (Float16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2025-08-14",
+      "com.github.apple.coremltools.source" : "torch==2.5.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 1501 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1501, 128]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 1]",
+        "name" : "input_1",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "AudioEncoder",
+    "method" : "predict"
+  }
+]

nvidia_parakeet-v3/AudioEncoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f95d33a2f1582d171ad8dacdb724d1d96a8566c958044c61419f1b90d307074b
+size 1219841984

nvidia_parakeet-v3/LICENSE_NOTICE.txt

@@ -0,0 +1,7 @@
+Argmax proprietary and confidential. Under NDA.
+
+Copyright 2024 Argmax, Inc. All rights reserved.
+
+Unauthorized access, copying, use, distribution, and or commercialization of this file, via any medium or means is strictly prohibited.
+
+Please contact Argmax for licensing information at [email protected].

nvidia_parakeet-v3/MelSpectrogram.mlmodelc/LICENSE_NOTICE.txt

@@ -0,0 +1,7 @@
+Argmax proprietary and confidential. Under NDA.
+
+Copyright 2024 Argmax, Inc. All rights reserved.
+
+Unauthorized access, copying, use, distribution, and or commercialization of this file, via any medium or means is strictly prohibited.
+
+Please contact Argmax for licensing information at [email protected].

nvidia_parakeet-v3/MelSpectrogram.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e7d1205a095f5ebc8372aec82b911979c3a010c37bd70fd2f546a21dbdce15c
+size 243

nvidia_parakeet-v3/MelSpectrogram.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f2a333c9134ecf776a755d626cb782de9fd65a4520bc18668432c21da75a625
+size 392

nvidia_parakeet-v3/MelSpectrogram.mlmodelc/metadata.json

@@ -0,0 +1,77 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Mixed (Float16, Palettized (6 bits))",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 1 × 1501 × 128)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1501, 128]",
+        "name" : "melspectrogram_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 2,
+      "Ios17.sqrt" : 1,
+      "Ios17.square" : 3,
+      "Ios17.transpose" : 1,
+      "Ios17.sub" : 2,
+      "Ios16.constexprLutToDense" : 3,
+      "Ios17.conv" : 2,
+      "Ios17.matmul" : 1,
+      "Ios17.log" : 1,
+      "Ios17.sliceByIndex" : 2,
+      "Ios17.add" : 3,
+      "Ios16.reduceMean" : 2,
+      "Ios17.realDiv" : 1,
+      "Ios17.expandDims" : 4,
+      "Ios17.squeeze" : 2,
+      "Ios17.reshape" : 2,
+      "Pad" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2025-08-14",
+      "com.github.apple.coremltools.source" : "torch==2.5.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 240000)",
+        "shortDescription" : "",
+        "shape" : "[240000]",
+        "name" : "audio",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "MelSpectrogram_6_bit",
+    "method" : "predict"
+  }
+]

nvidia_parakeet-v3/MelSpectrogram.mlmodelc/model.mil

@@ -0,0 +1,81 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}})]
+{
+    func main<ios17>(tensor<fp16, [240000]> audio) {
+            tensor<int32, [1]> var_8_begin_0 = const()[name = tensor<string, []>("op_8_begin_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> var_8_end_0 = const()[name = tensor<string, []>("op_8_end_0"), val = tensor<int32, [1]>([240000])];
+            tensor<bool, [1]> var_8_end_mask_0 = const()[name = tensor<string, []>("op_8_end_mask_0"), val = tensor<bool, [1]>([true])];
+            tensor<fp16, [239999]> var_8_cast_fp16 = slice_by_index(begin = var_8_begin_0, end = var_8_end_0, end_mask = var_8_end_mask_0, x = audio)[name = tensor<string, []>("op_8_cast_fp16")];
+            tensor<int32, [1]> var_13_begin_0 = const()[name = tensor<string, []>("op_13_begin_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> var_13_end_0 = const()[name = tensor<string, []>("op_13_end_0"), val = tensor<int32, [1]>([239999])];
+            tensor<bool, [1]> var_13_end_mask_0 = const()[name = tensor<string, []>("op_13_end_mask_0"), val = tensor<bool, [1]>([false])];
+            tensor<fp16, [239999]> var_13_cast_fp16 = slice_by_index(begin = var_13_begin_0, end = var_13_end_0, end_mask = var_13_end_mask_0, x = audio)[name = tensor<string, []>("op_13_cast_fp16")];
+            tensor<fp16, []> var_14_to_fp16 = const()[name = tensor<string, []>("op_14_to_fp16"), val = tensor<fp16, []>(0x1.f0cp-1)];
+            tensor<fp16, [239999]> var_15_cast_fp16 = mul(x = var_13_cast_fp16, y = var_14_to_fp16)[name = tensor<string, []>("op_15_cast_fp16")];
+            tensor<fp16, [239999]> input_1_cast_fp16 = sub(x = var_8_cast_fp16, y = var_15_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<int32, [2]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp16, []> const_0_to_fp16 = const()[name = tensor<string, []>("const_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [240000]> input_3_cast_fp16 = pad(constant_val = const_0_to_fp16, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [3]> var_30 = const()[name = tensor<string, []>("op_30"), val = tensor<int32, [3]>([1, 1, 240000])];
+            tensor<fp16, [1, 1, 240000]> input_5_cast_fp16 = reshape(shape = var_30, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<int32, [6]> input_7_pad_0 = const()[name = tensor<string, []>("input_7_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            tensor<string, []> input_7_mode_0 = const()[name = tensor<string, []>("input_7_mode_0"), val = tensor<string, []>("reflect")];
+            tensor<fp16, []> const_2_to_fp16 = const()[name = tensor<string, []>("const_2_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [1, 1, 240512]> input_7_cast_fp16 = pad(constant_val = const_2_to_fp16, mode = input_7_mode_0, pad = input_7_pad_0, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
+            tensor<int32, [1]> var_42 = const()[name = tensor<string, []>("op_42"), val = tensor<int32, [1]>([240512])];
+            tensor<fp16, [240512]> input_cast_fp16 = reshape(shape = var_42, x = input_7_cast_fp16)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 240512]> expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = input_cast_fp16)[name = tensor<string, []>("expand_dims_0_cast_fp16")];
+            tensor<int32, [1]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = tensor<string, []>("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 240512]> expand_dims_4_cast_fp16 = expand_dims(axes = expand_dims_4_axes_0, x = expand_dims_0_cast_fp16)[name = tensor<string, []>("expand_dims_4_cast_fp16")];
+            tensor<string, []> conv_0_pad_type_0 = const()[name = tensor<string, []>("conv_0_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = tensor<string, []>("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = tensor<string, []>("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_0_groups_0 = const()[name = tensor<string, []>("conv_0_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_1_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [98688]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98816))), name = tensor<string, []>("expand_dims_1_to_fp16_palettized"), shape = tensor<uint32, [3]>([257, 1, 512])];
+            tensor<fp16, [1, 257, 1501]> conv_0_cast_fp16 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1_to_fp16_palettized, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_0_cast_fp16")];
+            tensor<string, []> conv_1_pad_type_0 = const()[name = tensor<string, []>("conv_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = tensor<string, []>("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = tensor<string, []>("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_1_groups_0 = const()[name = tensor<string, []>("conv_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp16, [257, 1, 512]> expand_dims_2_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [98688]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(99008))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197760))), name = tensor<string, []>("expand_dims_2_to_fp16_palettized"), shape = tensor<uint32, [3]>([257, 1, 512])];
+            tensor<fp16, [1, 257, 1501]> conv_1_cast_fp16 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2_to_fp16_palettized, x = expand_dims_4_cast_fp16)[name = tensor<string, []>("conv_1_cast_fp16")];
+            tensor<int32, [1]> squeeze_0_axes_0 = const()[name = tensor<string, []>("squeeze_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [257, 1501]> squeeze_0_cast_fp16 = squeeze(axes = squeeze_0_axes_0, x = conv_0_cast_fp16)[name = tensor<string, []>("squeeze_0_cast_fp16")];
+            tensor<int32, [1]> squeeze_1_axes_0 = const()[name = tensor<string, []>("squeeze_1_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [257, 1501]> squeeze_1_cast_fp16 = squeeze(axes = squeeze_1_axes_0, x = conv_1_cast_fp16)[name = tensor<string, []>("squeeze_1_cast_fp16")];
+            tensor<fp16, [257, 1501]> square_1_cast_fp16 = square(x = squeeze_0_cast_fp16)[name = tensor<string, []>("square_1_cast_fp16")];
+            tensor<fp16, [257, 1501]> square_2_cast_fp16 = square(x = squeeze_1_cast_fp16)[name = tensor<string, []>("square_2_cast_fp16")];
+            tensor<fp16, [257, 1501]> add_1_cast_fp16 = add(x = square_1_cast_fp16, y = square_2_cast_fp16)[name = tensor<string, []>("add_1_cast_fp16")];
+            tensor<bool, []> mel_spec_1_transpose_x_0 = const()[name = tensor<string, []>("mel_spec_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> mel_spec_1_transpose_y_0 = const()[name = tensor<string, []>("mel_spec_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [128, 257]> mel_filters_to_fp16_palettized = constexpr_lut_to_dense()[indices = tensor<uint8, [24672]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(197952))), lut = tensor<fp16, [64]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(222720))), name = tensor<string, []>("mel_filters_to_fp16_palettized"), shape = tensor<uint32, [2]>([128, 257])];
+            tensor<fp16, [128, 1501]> mel_spec_1_cast_fp16 = matmul(transpose_x = mel_spec_1_transpose_x_0, transpose_y = mel_spec_1_transpose_y_0, x = mel_filters_to_fp16_palettized, y = add_1_cast_fp16)[name = tensor<string, []>("mel_spec_1_cast_fp16")];
+            tensor<fp16, []> var_56_to_fp16 = const()[name = tensor<string, []>("op_56_to_fp16"), val = tensor<fp16, []>(0x1p-24)];
+            tensor<fp16, [128, 1501]> mel_spec_3_cast_fp16 = add(x = mel_spec_1_cast_fp16, y = var_56_to_fp16)[name = tensor<string, []>("mel_spec_3_cast_fp16")];
+            tensor<fp32, []> mel_spec_5_epsilon_0 = const()[name = tensor<string, []>("mel_spec_5_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp16, [128, 1501]> mel_spec_5_cast_fp16 = log(epsilon = mel_spec_5_epsilon_0, x = mel_spec_3_cast_fp16)[name = tensor<string, []>("mel_spec_5_cast_fp16")];
+            tensor<int32, [1]> per_feature_mean_axes_0 = const()[name = tensor<string, []>("per_feature_mean_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> per_feature_mean_keep_dims_0 = const()[name = tensor<string, []>("per_feature_mean_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [128, 1]> per_feature_mean_cast_fp16 = reduce_mean(axes = per_feature_mean_axes_0, keep_dims = per_feature_mean_keep_dims_0, x = mel_spec_5_cast_fp16)[name = tensor<string, []>("per_feature_mean_cast_fp16")];
+            tensor<fp16, [128, 1501]> sub_0_cast_fp16 = sub(x = mel_spec_5_cast_fp16, y = per_feature_mean_cast_fp16)[name = tensor<string, []>("sub_0_cast_fp16")];
+            tensor<fp16, [128, 1501]> square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor<string, []>("square_0_cast_fp16")];
+            tensor<int32, [1]> reduce_mean_1_axes_0 = const()[name = tensor<string, []>("reduce_mean_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> reduce_mean_1_keep_dims_0 = const()[name = tensor<string, []>("reduce_mean_1_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp16, [128, 1]> reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor<string, []>("reduce_mean_1_cast_fp16")];
+            tensor<fp16, []> real_div_0_to_fp16 = const()[name = tensor<string, []>("real_div_0_to_fp16"), val = tensor<fp16, []>(0x1.004p+0)];
+            tensor<fp16, [128, 1]> mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor<string, []>("mul_0_cast_fp16")];
+            tensor<fp16, [128, 1]> sqrt_0_cast_fp16 = sqrt(x = mul_0_cast_fp16)[name = tensor<string, []>("sqrt_0_cast_fp16")];
+            tensor<fp16, []> var_70_to_fp16 = const()[name = tensor<string, []>("op_70_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [128, 1]> per_feature_std_cast_fp16 = add(x = sqrt_0_cast_fp16, y = var_70_to_fp16)[name = tensor<string, []>("per_feature_std_cast_fp16")];
+            tensor<fp16, [128, 1501]> mel_spec_cast_fp16 = real_div(x = sub_0_cast_fp16, y = per_feature_std_cast_fp16)[name = tensor<string, []>("mel_spec_cast_fp16")];
+            tensor<int32, [2]> var_75_perm_0 = const()[name = tensor<string, []>("op_75_perm_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<int32, [1]> var_77_axes_0 = const()[name = tensor<string, []>("op_77_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1501, 128]> var_75_cast_fp16 = transpose(perm = var_75_perm_0, x = mel_spec_cast_fp16)[name = tensor<string, []>("transpose_0")];
+            tensor<fp16, [1, 1501, 128]> var_77_cast_fp16 = expand_dims(axes = var_77_axes_0, x = var_75_cast_fp16)[name = tensor<string, []>("op_77_cast_fp16")];
+            tensor<int32, [1]> var_79_axes_0 = const()[name = tensor<string, []>("op_79_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 1501, 128]> melspectrogram_features = expand_dims(axes = var_79_axes_0, x = var_77_cast_fp16)[name = tensor<string, []>("op_79_cast_fp16")];
+        } -> (melspectrogram_features);
+}

nvidia_parakeet-v3/MelSpectrogram.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fdb3127d090c4856df950d5e6f059d1936ddd121a4f9da7362ed7c5ef7b189b3
+size 222912

nvidia_parakeet-v3/MultimodalLogits.mlmodelc/LICENSE_NOTICE.txt

@@ -0,0 +1,7 @@
+Argmax proprietary and confidential. Under NDA.
+
+Copyright 2024 Argmax, Inc. All rights reserved.
+
+Unauthorized access, copying, use, distribution, and or commercialization of this file, via any medium or means is strictly prohibited.
+
+Please contact Argmax for licensing information at [email protected].

nvidia_parakeet-v3/MultimodalLogits.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbb1fc22c1a4b16fff489b9424efcf0112cd0223ca427621aca45ce744f3ff8d
+size 243

nvidia_parakeet-v3/MultimodalLogits.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a0441fb4458b0ddb433752446048f5ff3fe681af731833b96a8fe8f80d0521b
+size 432

nvidia_parakeet-v3/MultimodalLogits.mlmodelc/metadata.json

@@ -0,0 +1,75 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 8198)",
+        "shortDescription" : "",
+        "shape" : "[1, 8198]",
+        "name" : "logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios16.softmax" : 1,
+      "Ios17.log" : 1,
+      "Ios17.linear" : 1,
+      "Ios17.add" : 1,
+      "Ios16.relu" : 1
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2025-08-14",
+      "com.github.apple.coremltools.source" : "torch==2.5.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 640]",
+        "name" : "encoder_output_projected",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 640]",
+        "name" : "decoder_output_projected",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "MultimodalLogits",
+    "method" : "predict"
+  }
+]

nvidia_parakeet-v3/MultimodalLogits.mlmodelc/model.mil

@@ -0,0 +1,15 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<fp16, [1, 640]> decoder_output_projected, tensor<fp16, [1, 640]> encoder_output_projected) {
+            tensor<fp16, [1, 640]> input_1_cast_fp16 = add(x = decoder_output_projected, y = encoder_output_projected)[name = tensor<string, []>("input_1_cast_fp16")];
+            tensor<fp16, [1, 640]> input_3_cast_fp16 = relu(x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<fp16, [8198, 640]> joint_net_1_weight_to_fp16 = const()[name = tensor<string, []>("joint_net_1_weight_to_fp16"), val = tensor<fp16, [8198, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [8198]> joint_net_1_bias_to_fp16 = const()[name = tensor<string, []>("joint_net_1_bias_to_fp16"), val = tensor<fp16, [8198]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10493568)))];
+            tensor<fp16, [1, 8198]> linear_0_cast_fp16 = linear(bias = joint_net_1_bias_to_fp16, weight = joint_net_1_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+            tensor<int32, []> var_11 = const()[name = tensor<string, []>("op_11"), val = tensor<int32, []>(-1)];
+            tensor<fp16, [1, 8198]> var_13_softmax_cast_fp16 = softmax(axis = var_11, x = linear_0_cast_fp16)[name = tensor<string, []>("op_13_softmax_cast_fp16")];
+            tensor<fp32, []> var_13_epsilon_0 = const()[name = tensor<string, []>("op_13_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp16, [1, 8198]> logits = log(epsilon = var_13_epsilon_0, x = var_13_softmax_cast_fp16)[name = tensor<string, []>("op_13_cast_fp16")];
+        } -> (logits);
+}

nvidia_parakeet-v3/MultimodalLogits.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d83381f8a19ac296033554a7cfec355063b8f3225129cb8bed013c92f6ab141f
+size 10510028

nvidia_parakeet-v3/TextDecoder.mlmodelc/LICENSE_NOTICE.txt

@@ -0,0 +1,7 @@
+Argmax proprietary and confidential. Under NDA.
+
+Copyright 2024 Argmax, Inc. All rights reserved.
+
+Unauthorized access, copying, use, distribution, and or commercialization of this file, via any medium or means is strictly prohibited.
+
+Please contact Argmax for licensing information at [email protected].

nvidia_parakeet-v3/TextDecoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91faa7cfe003e448303bbbec3fe4ae0cf18a80ec8d903f760867e905b09cc125
+size 243

nvidia_parakeet-v3/TextDecoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d337a032cea76091dc96a57d0c7a904c95190e54442a2fe336057ca0714378ce
+size 504

nvidia_parakeet-v3/TextDecoder.mlmodelc/metadata.json

@@ -0,0 +1,111 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float16",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[1, 640]",
+        "name" : "decoder_output_projected",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 640]",
+        "name" : "new_state_1",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 640]",
+        "name" : "new_state_2",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Select" : 1,
+      "Ios17.squeeze" : 7,
+      "Ios17.gather" : 1,
+      "Ios17.cast" : 3,
+      "Ios17.lstm" : 2,
+      "Split" : 2,
+      "Ios17.add" : 1,
+      "Ios17.linear" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Stack" : 2,
+      "Ios17.expandDims" : 3
+    },
+    "computePrecision" : "Mixed (Float16, Int16, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2025-08-14",
+      "com.github.apple.coremltools.source" : "torch==2.5.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "decoder_input_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 640]",
+        "name" : "state_1",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float16",
+        "formattedType" : "MultiArray (Float16 2 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 640]",
+        "name" : "state_2",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "TextDecoder",
+    "method" : "predict"
+  }
+]

nvidia_parakeet-v3/TextDecoder.mlmodelc/model.mil

@@ -0,0 +1,73 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3401.3.1"}, {"coremlc-version", "3401.4.1"}, {"coremltools-component-torch", "2.5.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<int32, [1]> decoder_input_ids, tensor<fp16, [2, 640]> state_1, tensor<fp16, [2, 640]> state_2) {
+            tensor<int32, []> input_1_batch_dims_0 = const()[name = tensor<string, []>("input_1_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> input_1_validate_indices_0 = const()[name = tensor<string, []>("input_1_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<fp16, [8193, 640]> prediction_embed_weight_to_fp16 = const()[name = tensor<string, []>("prediction_embed_weight_to_fp16"), val = tensor<fp16, [8193, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<string, []> decoder_input_ids_to_int16_dtype_0 = const()[name = tensor<string, []>("decoder_input_ids_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<string, []> cast_6_dtype_0 = const()[name = tensor<string, []>("cast_6_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
+            tensor<int16, [1]> decoder_input_ids_to_int16 = cast(dtype = decoder_input_ids_to_int16_dtype_0, x = decoder_input_ids)[name = tensor<string, []>("cast_9")];
+            tensor<int32, [1]> cast_6 = cast(dtype = cast_6_dtype_0, x = decoder_input_ids_to_int16)[name = tensor<string, []>("cast_8")];
+            tensor<bool, [1]> greater_equal_0 = greater_equal(x = cast_6, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
+            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(8193)];
+            tensor<int32, [1]> add_2 = add(x = cast_6, y = slice_by_index_0)[name = tensor<string, []>("add_2")];
+            tensor<int32, [1]> select_0 = select(a = cast_6, b = add_2, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
+            tensor<int32, []> input_1_cast_fp16_cast_uint16_axis_0 = const()[name = tensor<string, []>("input_1_cast_fp16_cast_uint16_axis_0"), val = tensor<int32, []>(0)];
+            tensor<string, []> select_0_to_int16_dtype_0 = const()[name = tensor<string, []>("select_0_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<int16, [1]> select_0_to_int16 = cast(dtype = select_0_to_int16_dtype_0, x = select_0)[name = tensor<string, []>("cast_7")];
+            tensor<fp16, [1, 640]> input_1_cast_fp16_cast_uint16_cast_uint16 = gather(axis = input_1_cast_fp16_cast_uint16_axis_0, batch_dims = input_1_batch_dims_0, indices = select_0_to_int16, validate_indices = input_1_validate_indices_0, x = prediction_embed_weight_to_fp16)[name = tensor<string, []>("input_1_cast_fp16_cast_uint16_cast_uint16")];
+            tensor<int32, [1]> input_3_axes_0 = const()[name = tensor<string, []>("input_3_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 1, 640]> input_3_cast_fp16 = expand_dims(axes = input_3_axes_0, x = input_1_cast_fp16_cast_uint16_cast_uint16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<int32, [1]> hx_1_axes_0 = const()[name = tensor<string, []>("hx_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 640]> hx_1_cast_fp16 = expand_dims(axes = hx_1_axes_0, x = state_1)[name = tensor<string, []>("hx_1_cast_fp16")];
+            tensor<int32, [1]> hx_axes_0 = const()[name = tensor<string, []>("hx_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 1, 640]> hx_cast_fp16 = expand_dims(axes = hx_axes_0, x = state_2)[name = tensor<string, []>("hx_cast_fp16")];
+            tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [1, 1, 640]> split_0_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_0_cast_fp16_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = hx_1_cast_fp16)[name = tensor<string, []>("split_0_cast_fp16")];
+            tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [1, 1, 640]> split_1_cast_fp16_0, tensor<fp16, [1, 1, 640]> split_1_cast_fp16_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = hx_cast_fp16)[name = tensor<string, []>("split_1_cast_fp16")];
+            tensor<int32, [1]> output_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("output_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> output_lstm_layer_0_lstm_h0_squeeze_cast_fp16 = squeeze(axes = output_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_0)[name = tensor<string, []>("output_lstm_layer_0_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> output_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("output_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> output_lstm_layer_0_lstm_c0_squeeze_cast_fp16 = squeeze(axes = output_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_0)[name = tensor<string, []>("output_lstm_layer_0_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> output_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("output_lstm_layer_0_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> output_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("output_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> output_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("output_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> output_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("output_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> output_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("output_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_1_to_fp16 = const()[name = tensor<string, []>("concat_1_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10487168)))];
+            tensor<fp16, [2560, 640]> concat_2_to_fp16 = const()[name = tensor<string, []>("concat_2_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13764032)))];
+            tensor<fp16, [2560]> concat_0_to_fp16 = const()[name = tensor<string, []>("concat_0_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17040896)))];
+            tensor<fp16, [1, 1, 640]> output_lstm_layer_0_cast_fp16_0, tensor<fp16, [1, 640]> output_lstm_layer_0_cast_fp16_1, tensor<fp16, [1, 640]> output_lstm_layer_0_cast_fp16_2 = lstm(activation = output_lstm_layer_0_activation_0, bias = concat_0_to_fp16, cell_activation = output_lstm_layer_0_cell_activation_0, direction = output_lstm_layer_0_direction_0, initial_c = output_lstm_layer_0_lstm_c0_squeeze_cast_fp16, initial_h = output_lstm_layer_0_lstm_h0_squeeze_cast_fp16, output_sequence = output_lstm_layer_0_output_sequence_0, recurrent_activation = output_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2_to_fp16, weight_ih = concat_1_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("output_lstm_layer_0_cast_fp16")];
+            tensor<int32, [1]> output_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("output_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> output_lstm_h0_squeeze_cast_fp16 = squeeze(axes = output_lstm_h0_squeeze_axes_0, x = split_0_cast_fp16_1)[name = tensor<string, []>("output_lstm_h0_squeeze_cast_fp16")];
+            tensor<int32, [1]> output_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("output_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp16, [1, 640]> output_lstm_c0_squeeze_cast_fp16 = squeeze(axes = output_lstm_c0_squeeze_axes_0, x = split_1_cast_fp16_1)[name = tensor<string, []>("output_lstm_c0_squeeze_cast_fp16")];
+            tensor<string, []> output_direction_0 = const()[name = tensor<string, []>("output_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> output_output_sequence_0 = const()[name = tensor<string, []>("output_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> output_recurrent_activation_0 = const()[name = tensor<string, []>("output_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> output_cell_activation_0 = const()[name = tensor<string, []>("output_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> output_activation_0 = const()[name = tensor<string, []>("output_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp16, [2560, 640]> concat_4_to_fp16 = const()[name = tensor<string, []>("concat_4_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17046080)))];
+            tensor<fp16, [2560, 640]> concat_5_to_fp16 = const()[name = tensor<string, []>("concat_5_to_fp16"), val = tensor<fp16, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20322944)))];
+            tensor<fp16, [2560]> concat_3_to_fp16 = const()[name = tensor<string, []>("concat_3_to_fp16"), val = tensor<fp16, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23599808)))];
+            tensor<fp16, [1, 1, 640]> output_cast_fp16_0, tensor<fp16, [1, 640]> output_cast_fp16_1, tensor<fp16, [1, 640]> output_cast_fp16_2 = lstm(activation = output_activation_0, bias = concat_3_to_fp16, cell_activation = output_cell_activation_0, direction = output_direction_0, initial_c = output_lstm_c0_squeeze_cast_fp16, initial_h = output_lstm_h0_squeeze_cast_fp16, output_sequence = output_output_sequence_0, recurrent_activation = output_recurrent_activation_0, weight_hh = concat_5_to_fp16, weight_ih = concat_4_to_fp16, x = output_lstm_layer_0_cast_fp16_0)[name = tensor<string, []>("output_cast_fp16")];
+            tensor<int32, []> var_32_axis_0 = const()[name = tensor<string, []>("op_32_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> var_32_cast_fp16 = stack(axis = var_32_axis_0, values = (output_lstm_layer_0_cast_fp16_1, output_cast_fp16_1))[name = tensor<string, []>("op_32_cast_fp16")];
+            tensor<int32, []> var_33_axis_0 = const()[name = tensor<string, []>("op_33_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp16, [2, 1, 640]> var_33_cast_fp16 = stack(axis = var_33_axis_0, values = (output_lstm_layer_0_cast_fp16_2, output_cast_fp16_2))[name = tensor<string, []>("op_33_cast_fp16")];
+            tensor<int32, [1]> input_axes_0 = const()[name = tensor<string, []>("input_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [1, 640]> input_cast_fp16 = squeeze(axes = input_axes_0, x = output_cast_fp16_0)[name = tensor<string, []>("input_cast_fp16")];
+            tensor<int32, [1]> var_35_axes_0 = const()[name = tensor<string, []>("op_35_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 640]> new_state_1 = squeeze(axes = var_35_axes_0, x = var_32_cast_fp16)[name = tensor<string, []>("op_35_cast_fp16")];
+            tensor<int32, [1]> var_36_axes_0 = const()[name = tensor<string, []>("op_36_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp16, [2, 640]> new_state_2 = squeeze(axes = var_36_axes_0, x = var_33_cast_fp16)[name = tensor<string, []>("op_36_cast_fp16")];
+            tensor<fp16, [640, 640]> joint_projection_weight_to_fp16 = const()[name = tensor<string, []>("joint_projection_weight_to_fp16"), val = tensor<fp16, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(23604992)))];
+            tensor<fp16, [640]> joint_projection_bias_to_fp16 = const()[name = tensor<string, []>("joint_projection_bias_to_fp16"), val = tensor<fp16, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24424256)))];
+            tensor<fp16, [1, 640]> decoder_output_projected = linear(bias = joint_projection_bias_to_fp16, weight = joint_projection_weight_to_fp16, x = input_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
+        } -> (decoder_output_projected, new_state_1, new_state_2);
+}

nvidia_parakeet-v3/TextDecoder.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:36cd656ebf1105ee5aa85d1e55c00f25669327aef8b39346333a19260ca78018
+size 24425600