Skip to content

Deckard v0.6.0 — Apple Neural Engine

Choose a tag to compare

View all tags
@sgoedecke sgoedecke released this 13 Sep 08:40
· 17 commits to main since this release
v0.6.0
f835d65

Default-priority Apple Neural Engine

Deckard now runs Gradient through public Core ML CPU/Neural Engine APIs instead of MLX/Metal. There is no GPU fallback or private ANE runtime hook.

  • Includes the pinned FP16 model and tokenizer in the release archive (about 581.5 MiB compressed). No separate Hugging Face download, Python, or Xcode is required.
  • Keeps the model resident and caches local compilation across host lifetimes. The first model use can take roughly 26 seconds on an M4; subsequent resident classifications averaged about 188 ms in development measurements. Other Macs may differ.
  • Preserves the 50-word minimum, windowing, saved preferences, and default 0.97 cutoff. All 52 frozen reference windows retained the same default-cutoff decisions.
  • Supports upgrading existing Deckard MLX installations in place.

Install or upgrade

curl --proto '=https' --tlsv1.2 -fsSL https://github.com/sgoedecke/deckard/releases/download/v0.6.0/install.sh | bash

After upgrading, click Reload on Deckard at chrome://extensions. Apple Silicon and macOS 15 or newer are required. The native executable is ad-hoc signed, not notarized.

The FP16 model is derived from decoded canonical q4 weights; it does not recover the original FP32 weights. Scores remain experimental and are not proof of AI authorship. Model attribution and MIT notices are bundled; see model provenance.

SHA256SUMS covers the archive and installer; the installer also embeds the archive checksum.

Assets 5
Loading