Download Latest Version vMLX 1.6.84 source code.zip (2.1 MB)
Home / v1.6.81
Name Modified Size InfoDownloads / Week
Parent folder
release-info.json 2026-10-21 2.0 kB
vMLX-1.6.81-sequoia-arm64.dmg 2026-10-21 578.8 MB
vMLX-1.6.81-sequoia-arm64.dmg.blockmap 2026-10-21 606.8 kB
vMLX-1.6.81-tahoe-arm64.dmg 2026-10-21 604.3 MB
vMLX-1.6.81-tahoe-arm64.dmg.blockmap 2026-10-21 631.9 kB
README.md 2026-10-18 2.2 kB
vMLX 1.6.81 source code.tar.gz 2026-10-18 2.1 MB
vMLX 1.6.81 source code.zip 2026-10-18 2.1 MB
Totals: 8 Items   1.2 GB 0

vMLX 1.6.81

  • Let native MTP, DFlash2 and DSpark choose draft work adaptively within their verified limits. Treat bundle and tuning depths as starting points, with measured ordinary-decoding fallback and re-entry.
  • Add explicit Adaptive LiLiCorr controls for compatible Nemotron Bolt bundles. Preserve the existing off-by-default behavior and saved session choice.
  • Correct native verification at long-context attention dispatch boundaries and Qwen DFlash2 verification across wider blocks. Preserve causal presence-penalty arithmetic and rejected-token rollback.
  • Honor model-declared thinking and non-thinking sampling presets across the app and all four chat APIs, including embedded JANG configuration. Keep explicit request overrides.
  • Improve Ornith 1.5 loading, prefill and long-context attention, including split expert layouts, hidden-state normalization and idempotent MTP patch installation.
  • Improve Qwen 27B dense multi-row decoding and keep MiMo verification on its qualified MLX kernels. Throughput gains depend on the bundle, sampling, context and workload; adaptive operation does not guarantee faster requests or improved prefill.
  • Preserve speculative SSD parent chains and native media positions across restores. Reduce duplicate hybrid-cache payload work while retaining background persistence and safe tool continuations.
  • Report runtime-selected draft widths, scoped native MTP acceptance, mixed weight precision and observed GDN dispatch accurately. Keep cached prompt throughput distinct from newly computed prefill.
  • Explain suppressed model tool markup without emitting premature success. Preserve consumed token/cache usage on API errors and completed empty assistant boundaries in subsequent chat history.
  • Make generic external drafting adaptive and verify sequentially to preserve ordinary target arithmetic. This compatibility path can be slower than ordinary decoding and does not support persistent prefix caching.
  • Preserve small file sizes in built-in directory listings.

Downloads: Tahoe build for macOS 26 and later; Sequoia compatibility build for supported earlier macOS versions. Both require Apple Silicon.

Source: README.md, updated 2026-10-18