The previous files used make_quantizer's MovingAverageMinMaxObserver default,
whose moving average trails the real activation range. Measured on an S26 Ultra
over 500 ImageNetV2 images: 70.00% top-1 before, 71.80% after, against 72.20%
for fp32. Quantization loss drops from 2.20 points to 0.40 at the same size.
Only the v81 file is device-validated; the other four are the same code path and
calibration compiled for Hexagon versions we have no hardware for.
Add QNN a16w8 exports, one per Hexagon version (v69-v81)
quantized duplicated precisions.json, which is authoritative for the
quantized/non-quantized partition. default expressed a per-repository
recommendation nothing consumed: the library resolves its own per-platform
default. Both are gone from the schema.
Adds library_name to the model card, and pipeline_tag where it was missing.
Brings the config and card licenses into agreement, and repairs the variant
declarations that did not match the files actually published here.
Core ML: drop the fp32 variant, fp16 is equivalent in task terms
fp32 is both slower and larger than fp16 on device, and the ANE will not accept fp32 at all, so an fp32 Core ML build is structurally excluded from the accelerator that makes Core ML worth using.
Equivalence was gated with a task metric rather than raw tensor distance: top-1 agreement 100% across 8 real images. Raw relative L2 on logits reads ~1% for the segmentation models, but that is logit noise which does not change the predicted class per pixel.
Models where fp16 and fp32 genuinely diverge (ssdlite320 86% box agreement, fast-sam 93%) keep both variants.
v0.9.0 and below are untouched.
The previous removal commit said earlier tags still contain the MLX artifacts. That is wrong: MLX for this model was only ever published on main / v0.10.0, and no tag at v0.9.0 or below ever carried it. The files remain recoverable from git history. v0.9.0 and below are, and have been, untouched.
Remove MLX artifacts: 7.5-19.6x slower than Core ML on device
MLX measured substantially slower than Core ML on an iPhone 16 for every convolutional vision model tested (efficientnet-v2-s 19.6x, yolo26-s 16.0x, yolo26-seg-s 7.5x, yolo26-pose-n 10.0x). MLX weight quantization also cannot shrink a conv model, since it only rewrites nn.Linear/embedding modules.
The .pte and config.json are removed; each mlx/ dir keeps a NOTES.md with the measurements. The exporter is retained in export-scripts so this can be re-tested against a future ExecuTorch/MLX release.
Earlier tags (v0.9.0 and below) are untouched and still contain the MLX artifacts.
