mlx2coreml 0.1.1

Experimental MLX to Core ML translation pipeline.

mlx2coreml

mlx2coreml is a MLX -> Core ML translation pipeline.

It captures MLX graphs and converts them to to Core ML models (.mlpackage / .mlmodelc), suitable for on-device deployment through Xcode.

mlx2coreml is still experimental and may not work with all models.

Quick Start

pip install -U mlx2coreml

Examples

Converting MLX LM models to CoreML

python -m mlx2coreml.convert_mlx_lm \
  --model-id mlx-community/Qwen3-0.6B-bf16 \
  --output qwen3_0_6b_coreml \
  --seq-len 128 \
  --deployment-target iOS18 \
  --convert-to mlprogram \
  --compute-precision fp16 \
  --compute-units all

Converting MLX Audio Smart Turn V3 to CoreML

python -m mlx2coreml.convert_mlx_audio \
  --model-id mlx-community/smart-turn-v3 \
  --output smart_turn_v3_coreml \
  --audio-seconds 8 \
  --deployment-target iOS18 \
  --convert-to mlprogram \
  --compute-precision fp16 \
  --compute-units all

Note: This captures the model's input_features graph, which are preprocessed log-mel features; raw waveform preprocessing stays outside the converted Core ML model.

Direct conversion from Python

For arbitrary models, use the conversion API directly:

import numpy as np
import mlx.core as mx

from mlx_lm import load

import mlx2coreml as m2c

# Load an MLX LM model

model_id = "mlx-community/Qwen3-0.6B-bf16"
mlx_model, tokenizer = load(model_id, lazy=False)

# Build sample inputs for conversion

seq_len = 128
prompt = "hello"
input_ids = mx.array(tokenizer.encode(prompt), dtype=mx.int32)[None, ...]
inputs = {"input_ids": input_ids}

# Convert the model to a CoreML .mlprogram

config = m2c.ConversionConfig(
    deployment_target="iOS18",
    convert_to="mlprogram",
    compute_precision="fp16",
    flex_input_lens=[1, seq_len],
    flex_input_names={"input_ids"},
)

coreml_model = m2c.convert_mlx_to_coreml(
    mlx_model,
    inputs,
    config=config,
    capture_function=lambda input_ids: mlx_model(input_ids),
)
model_path = "model.mlpackage"
coreml_model.model.save(model_path)

# Optional: compile the model to a CoreML .mlmodelc

compiled_path = m2c.compile_mlmodel(model_path, "model.mlmodelc")

Note that for some models, you need to prepare the input features if there's additional preprocessing required:

import numpy as np

from mlx_audio.vad import load

import mlx2coreml as m2c

model = load("mlx-community/smart-turn-v3")
features = model.prepare_input_features(np.zeros((16000,), dtype=np.float32), sample_rate=16000)

coreml_model = m2c.convert_mlx_to_coreml(
    model,
    {"input_features": features},
    config=m2c.ConversionConfig(
        deployment_target="iOS18",
        convert_to="mlprogram",
        compute_precision="fp16",
    ),
    capture_function=lambda input_features: model(input_features),
)

model_path = "model.mlpackage"
coreml_model.model.save(model_path)

# Optional: compile the model to a CoreML .mlmodelc

compiled_path = m2c.compile_mlmodel(model_path, "model.mlmodelc")

Docs

• docs/ops_status.md: Supported op status