vMLX 1.6.22

vMLX 1.6.22

Performance

• DeepSeek V4 Flash decodes ~11% faster at every context length. The MLX
  cache ceiling was a fixed 8 GB, which made MLX free and re-allocate decode
  graph buffers it could have reused. It now scales with installed memory
  (16 GB → 4 GB, 64 GB → 11.5 GB, 128 GB → 23 GB, capped at 24 GB). Peak
  process memory is unchanged, because this bounds reclaimable cache rather
  than reserving it.

  ceiling   6835 tok   7715 tok   9915 tok
     8 GB     17.6       17.5       10.5
    24 GB     19.5       19.4       11.4
  

Fixed

• Tool markup carrying a non-string name, such as {"name": 12345}, no longer
  produces a tool call whose function name is empty. Affected the Hermes, Qwen,
  Nous and Gemma 4 parsers; such markup is now treated as ordinary content.
• Tool-choice enforcement errors report the caller's own tool_choice instead
  of always naming 'required', so a request that pinned a specific function
  can tell which setting failed.
• DeepSeek V4 Flash delta-cache restore reads its block size and anchor interval
  from the cache records rather than assuming fixed values.

Downloads

Tahoe (macOS 26) is the recommended download. Use the Sequoia build for
macOS 14.5 / 15. Both DMGs are signed with Developer ID and notarized by Apple.

tahoe    sha256 8c1645dfad3907db9df5682ec4d1293d09fdf87a0331788feb30e92248303155
sequoia  sha256 c7ef8e36f51ef4a99d9d9008da1a3885845cf9ec78c3e1a9c8d0ed8fc4ba5859