Skip to content

vMLX 1.6.18

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 27 Jul 19:39
v1.6.18
fe7ccc5

vMLX 1.6.18 for macOS

This checkpoint updates the Electron app and bundled Python engine from source commit ded3f9f02e324f9aa5df4f3a331b7f7d4df40d64.

What changed

  • Auto reasoning follows the selected model bundle's native policy, while reasoning markers remain separated from visible answers.
  • Tool availability and authorization stay bound to the correct chat turn, with improved multi-tool, cancellation, LFM, media, and tool-result continuations.
  • MiniMax request-schema XML tool arguments are parsed without leaking native tool markup into visible answers.
  • Fresh sessions inherit sampling, parser, output, template, and native MTP settings from the selected model bundle; explicit saved session overrides remain distinct.
  • Cache status now reports actual prefix reuse, RAM/SSD blocks, uncached suffixes, reconstruction, fallback prefill, and total SSD usage more clearly.
  • SSD cache storage gains a shared size budget, coordinated LRU eviction, and publication outside the request-completion path.
  • Paged-cache prefix reuse now rejects recycled block IDs unless the full ordered cache chain still matches and is pinned atomically.
  • Laguna receives safer mixed-SWA cache handling, preserved tool history, incomplete-terminal handling, and affine-JANG JIT policy updates.
  • Chat rendering better preserves KaTeX math, currency, Markdown, HTML, XML, code fences, and in-progress reasoning text.
  • Localized UI state is more stable, obsolete migration/update notices are removed, local attachment playback is allowed, and Ollama stream failures remain visible as failures.
  • The bundled runtime requires JANG 2.5.34.

Choose the Sequoia or Tahoe arm64 DMG that matches your macOS release.

This is a public checkpoint rather than exhaustive closure of every model-family, modality, long-context, cache-eviction, gateway/LAN recovery, or JANG conversion scenario.

Thanks to GitHub @Hornsan1 for reporting many of the issues addressed in this release.

Assets 6
Loading