An .mlmodel file is one Protocol
Buffers message, CoreML.Specification.Model, from
Model.proto
in Apple’s coremltools repository (with FeatureTypes.proto), read
2026-11-12. The number before each field is its field number.
Model
| Field | Name | Meaning |
|---|---|---|
| 1 | specificationVersion | The oldest Core ML that can run it; see versions. |
| 2 | description | ModelDescription. |
| 10 | isUpdatable | The model can be trained further on the device. |
| 200 and up | Type (oneof) | Exactly one model body, below. |
| Field | Model type | Field | Model type |
|---|---|---|---|
| 200 | pipelineClassifier | 555 | customModel |
| 201 | pipelineRegressor | 556 | linkedModel |
| 202 | pipeline | 560 | classConfidenceThresholding |
| 300 | glmRegressor | 600 | oneHotEncoder |
| 301 | supportVectorRegressor | 601 | imputer |
| 302 | treeEnsembleRegressor | 602 | featureVectorizer |
| 303 | neuralNetworkRegressor | 603 | dictVectorizer |
| 304 | bayesianProbitRegressor | 604 | scaler |
| 400 | glmClassifier | 606 | categoricalMapping |
| 401 | supportVectorClassifier | 607 | normalizer |
| 402 | treeEnsembleClassifier | 609 | arrayFeatureExtractor |
| 403 | neuralNetworkClassifier | 610 | nonMaximumSuppression |
| 404 | kNearestNeighborsClassifier | 900 | identity |
| 500 | neuralNetwork | 2000–2006 | textClassifier, wordTagger, visionFeaturePrint, soundAnalysisPreprocessing, gazetteer, wordEmbedding, audioFeaturePrint |
| 501 | itemSimilarityRecommender | 3000 | serializedModel |
| 502 | mlProgram |
ModelDescription and Metadata
| Field | Name | Meaning |
|---|---|---|
| 1 | input | Repeated FeatureDescription: 1 name, 2 shortDescription, 3 type (int64, double, string, image, multi-array, dictionary, sequence or state). |
| 10 | output | Repeated FeatureDescription. |
| 11 | predictedFeatureName | For classifiers: the output holding the class. |
| 12 | predictedProbabilitiesName | For classifiers: the output holding the probabilities. |
| 13 | state | Stateful models’ state features. |
| 20, 21 | functions, defaultFunctionName | Models with several functions. |
| 50 | trainingInput | Inputs for on-device training. |
| 100 | metadata | Metadata: 1 shortDescription, 2 versionString, 3 author, 4 license, 100 userDefined (string map). |
Specification versions
| Version | Core ML | iOS |
|---|---|---|
| 1 | 1 | 11 |
| 2 | 1.2 | 11.2 |
| 3 | 2 | 12 |
| 4 | 3 | 13 |
| 5 | 4 | 14 |
| 6 | 5 | 15 |
| 7 | 6 | 16 |
| 8 | 7 | 17 |
| 9 | 8 | 18 |
| 10 | 9 | 26 |
The matching macOS, tvOS, watchOS and visionOS releases are listed in
Model.proto. The ML Program model type (mlProgram) arrived
with specification version 6, and updatable models with version 4.