onnx2coreml 1.0.0

Convert ONNX models to Apple Core ML — .mlpackage (ML Program) and .mlmodel (NeuralNetwork) — built on coremltools' MIL builder, with numerical-parity verification.

onnx2coreml

Convert ONNX models to Apple Core ML — .mlpackage (ML Program / MIL, primary) and .mlmodel (NeuralNetwork, secondary) — built on coremltools' MIL builder, with numerical-parity verification against ONNX Runtime.

• ML Program first for modern on-device performance (fp16 and Apple Neural Engine), with full NeuralNetwork support for older-OS reach.
• Maintainable as Core ML evolves: all coremltools coupling is isolated behind two facades, so spec/format changes touch two files, not the operator lowerings.
• Verified, not assumed: every operator has numerical-parity tests run against the live Core ML runtime on macOS.

Status: v1.0.0. Design: docs/specs/.

Install

uv pip install -e ".[dev,verify,test]"   # development
# or, once published:  uv pip install onnx2coreml

Requires Python 3.11–3.13 and macOS for prediction/verification (model building works anywhere coremltools installs).

Quickstart

import onnx2coreml as o2c

# ML Program (.mlpackage) — the default, fp16, iOS17+
mlmodel = o2c.convert("model.onnx", format="mlpackage", minimum_deployment_target="iOS17")
mlmodel.save("model.mlpackage")

# NeuralNetwork (.mlmodel)
o2c.convert("model.onnx", format="mlmodel").save("model.mlmodel")

# Check coverage before converting
report = o2c.analyze("model.onnx")
print(report.convertible, report.unsupported)

# Numerical parity vs ONNX Runtime
print(o2c.verify("model.onnx", "model.mlpackage"))

convert() returns a coremltools MLModel. Options: format, minimum_deployment_target, compute_precision (fp16/fp32), compute_units, fuse, verify.

CLI

onnx2coreml convert model.onnx -o model.mlpackage --format mlpackage --verify
onnx2coreml inspect model.onnx          # op histogram + per-format convertibility
onnx2coreml verify  model.onnx model.mlpackage
onnx2coreml schema                      # version, supported op count, error codes

All subcommands accept --json.

Supported operators (105 ops)

Family	Operators
Elementwise / math	Add, Sub, Mul, Div, Pow, Sqrt, Exp, Log, Abs, Erf, Reciprocal, Neg, Floor, Ceil, Round, Sign, Sin, Cos, Tan, Asin, Acos, Atan, Sinh, Cosh, Atanh, Min, Max, Clip, Mod, Equal, Greater, Less, GreaterOrEqual, LessOrEqual, And, Or, Xor, Not, IsNaN, Where
Activations	Relu, LeakyRelu, PRelu, Sigmoid, Tanh, Gelu, Softmax, LogSoftmax, Softplus, Elu, HardSigmoid, HardSwish
Convolution / pooling	Conv, ConvTranspose, MaxPool, AveragePool, GlobalAveragePool, GlobalMaxPool
Normalization	BatchNormalization, LayerNormalization, InstanceNormalization, GroupNormalization
Linear	MatMul, Gemm, Inverse (com.microsoft)
Shape / movement	Reshape, Transpose, Flatten, Squeeze, Unsqueeze, Concat, Split, Pad, Cast, Identity, DepthToSpace, GridSample, Resize, Upsample
Indexing / gather-scatter	Gather, GatherND, GatherElements, ScatterND, ScatterElements, NonZero, Slice, Expand, Tile, Shape, ConstantOfShape
Reduction	ReduceMean, ReduceSum, ReduceMax, ReduceMin, ReduceProd, ReduceL1, ReduceL2, ReduceLogSum, ReduceLogSumExp, ReduceSumSquare, ArgMax, ArgMin, TopK
Recurrent	LSTM
Attention	ScaledDotProductAttention (produced by the attention fusion pass)

onnx2coreml inspect <model> reports exactly which ops a given model needs and whether each is covered. Unsupported ops raise a single aggregated UnsupportedOpError; register a custom lowering with Converter().register("OpType") as an escape hatch.

How it works

ONNX → load → passes (opset/shape/fold/cleanup) → fusion (attention→SDPA) →
coverage-gate → lower to MIL (mb.*) → coremltools backend → .mlpackage / .mlmodel → verify

See docs/specs/2026-06-15-onnx2coreml-design.md for the full design and AGENTS.md for the developer guide.

Known limitations (v1.0.0)

• Fixed input shapes.
• A full all-axes reduction with keepdims=0 yields shape (1,) (Core ML has no rank-0 scalar).
• Resize linear+half_pixel via explicit sizes is .mlpackage-only (the NeuralNetwork backend lacks the sampling mode); the scales path covers it on both formats.

License

BSD-3-Clause.