New Core ML model package:
- beats_as2m_10s BEATs iter3+ fine-tuned on AudioSet-2M (cpt1)
Audio representation / sound-event tagger over a fixed 10-second 16 kHz
window: input `waveform`, output `probs` over the AudioSet ontology
(527 classes). FP16, exported with coremltools 9.0 from TorchScript.
BEATs is MIT-licensed (microsoft/unilm). Document it in README.md (I/O,
shapes, precision, conversion toolchain, SHA-256 checksum) and add a BEATs
component entry to THIRD_PARTY_LICENSES.md; the bundle now carries two MIT
models (Silero VAD, BEATs) alongside the Apache-2.0 ones.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
New Core ML model packages:
- qwen3_5_mlx_4b_vision_vid_768x480_w8 Qwen3.5-4B vision tower @ 768x480
- gemma4_mlx_e4b_vision_768x480_w8 Gemma 4 (e4b) vision tower @ 768x480
Re-convert the existing Qwen3.5-4B vision tower @ 512x320 to the
tensor_video_rgb variant: inputs are now two planar RGB fp16 tensors
(1,3,320,512) in 0..255 with x/127.5-1 normalisation baked into the graph,
replacing the previous BGRA CVPixelBuffer image inputs. The new 768x480
Qwen build shares that I/O contract.
All new/updated towers are Apache-2.0 (Gemma 4 weights included -- unlike
earlier Gemma releases). Update README.md (I/O, shapes, precision,
conversion toolchain, SHA-256 checksums) and add a Gemma 4 upstream
component entry to THIRD_PARTY_LICENSES.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Optimize yolox_l_1024x640_w8 for full-ANE execution
Replace the YOLOX-L Core ML package with a re-converted build that keeps
every operation on the Apple Neural Engine, eliminating the CPU-bound ops
in the previous conversion. Functionally equivalent (same YOLOX-L model,
1024x640 input, 8-bit weights); same archive layout and I/O contract.
Update README.md with the new SHA-256 checksum and conversion toolchain
(torch 2.12.1).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add VPIPE supplementary CoreML models with docs and licenses
Release three pre-converted Core ML model packages consumed by the VPIPE
runtime, each shipped as a Git LFS .tar of an .mlpackage:
- qwen3_5_mlx_4b_vision_vid_512x320_w8 Qwen3.5-4B vision tower (Apache-2.0)
- yolox_l_1024x640_w8 YOLOX-L object detector (Apache-2.0)
- silero-vad-unified-v6 Silero VAD v6, unified (MIT)
Document each model in README.md (task, I/O, precision, upstream source,
conversion toolchain, SHA-256 checksums) and record per-model attribution
with verbatim license texts in THIRD_PARTY_LICENSES.md.
The bundle mixes two licenses, so the README front matter declares
`license: other` (apache-2.0 and mit) rather than a single identifier.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
