The shipped 560/1120/2240 ms encoders with the conv-cache slice of a concat rewritten, so M2 and M3 Neural Engines compile them instead of running the encoder on the CPU. `weight.bin`, `coremldata.bin` and the inputs/outputs are unchanged and the outputs are byte-identical; only `model.mil` differs, made by `patch_mlmodelc.py`. One copy per tier serves both `latin` and `multilingual`, whose encoders are identical. 4480 ms has no such slice.
Kept out of `<model>/<tier>ms/` because FluidAudio releases download every file in that folder.
Issue: https://github.com/FluidInference/FluidAudio/issues/1141 (M2 Pro) and https://github.com/FluidInference/FluidAudio/issues/1116 (M3 Pro)
Verified on M2 Pro: https://github.com/FluidInference/FluidAudio/issues/1141#issuecomment-6757399739
Test builds (same bytes): https://huggingface.co/FluidInference/Nemotron-3.5-ASR-Streaming-Multilingual-0.6b-CoreML/discussions/3
FluidAudio PR: linked from the model card Changelog.