Skip to content

vMLX 1.6.42

Choose a tag to compare

@jjang-ai jjang-ai released this 27 Aug 11:53
· 259 commits to main since this release

vMLX 1.6.42 adds an owned Qwen3.8 Flash Next runtime with architecture-native SSD caching, multimodal input, preserved n-gram state, and native adaptive MTP.

Added

  • Qwen3.8 Flash Next support for its hybrid QSA/Gated DeltaNet architecture, SSD-backed PLE bigram/trigram tables, image and video inputs, native reasoning tiers, and in-model MTP head.
  • Electron settings for detected native MTP mode and adaptive or fixed D1-D3 depth policy, with persisted launch-argument parity.

Performance

  • Fused Gated DeltaNet projections and hyper-connection execution while preserving the checkpoint's mixed-precision JANG contract.
  • Request-local MTP rollback/commit state and adaptive depth probing based on confirmed tokens per wall second.
  • File-backed, row-addressed PLE tables that avoid materializing the full n-gram table in unified memory.

Fixed

  • Full-precision SSD prefix blocks now preserve QSA K/V/indexer state with typed Gated DeltaNet and PLE companion state across divergent suffixes and process restarts, without retained paged-cache RAM.
  • Mixed image/video histories preserve modality order, processor scaling, embeddings, cache keys, and request-local MTP state.
  • Native MTP enforces the complete greedy sampler contract; the explicitly enabled sampled path uses stochastic speculative acceptance.
  • Tool-bearing requests remain capped at D1 so adaptive probing cannot cross an uncommitted tool-call boundary.

Downloads

  • macOS Tahoe / macOS 26: vMLX-1.6.42-tahoe-arm64.dmg — recommended and default.
  • macOS Sequoia: vMLX-1.6.42-sequoia-arm64.dmg — compatibility build.

Both DMGs contain vMLX 1.6.42, MLX 0.32.2, and JANG 2.5.46 from pinned source revisions. Both were Developer ID signed, independently accepted by Apple notarization, stapled, Gatekeeper verified, and checked from their final mounted payloads. The final Tahoe app was installed on macOS 26 and visibly displayed Version 1.6.42 from its packaged app.asar; no model was loaded during the installed release smoke. The source-matched Qwen3.8 runtime evidence was retained because this frozen release commit changes version metadata only.