Skip to content

vMLX 1.6.50

Choose a tag to compare

@jjang-ai jjang-ai released this 01 Sep 04:42

vMLX 1.6.50

Changes

  • Pin PyPI publisher to executable commit (f36b1305)
  • Publish vMLX 1.6.49 updater manifest (4337b25f)
  • Run website deployment after release recovery (287cbd6d)
  • Enable qualified GLM KDA decode fusions by default (eb0ddc5e)
  • Split tight-memory GLM prefills before Metal OOM (b632b457)
  • Honor GLM prefill chunks at SSM boundaries (6ab04718)
  • Persist parsed tool turns before publishing terminals (51ead4c4)
  • Reconstruct stable tool contracts across turns (50ca236f)
  • Drain terminalizing tool requests without aborting (b6aae528)
  • Keep mixed-layout Qwen4 MoE decode atomic (9eba14de)
  • Prepare vMLX 1.6.50 (c29aeb27)

Runtime provenance

  • vMLX source: c29aeb27cf6c106ea21508b844a6348b97ccc412
  • JANG runtime: f3c79081c3d99e8a26fa28444934f7567acb00c6 (jang 2.5.47)
  • Tahoe/macOS 26 is the default build; Sequoia is the compatibility build.
  • Both DMGs are Developer ID signed, independently notarized, stapled, and Gatekeeper checked by the candidate workflow.

Downloads

  • Tahoe: vMLX-1.6.50-tahoe-arm64.dmg
    • SHA-256: e71f845db196656d046102180af757b95b70b644949fd362afda7a660688d34e
    • Size: 538306982 bytes
  • Sequoia: vMLX-1.6.50-sequoia-arm64.dmg
    • SHA-256: c8034d664bf80cac27f4a8cf20b6d8799850bf1cd81d752b533156d15d5e04e6
    • Size: 516197292 bytes

Python distributions

  • Wheel: vmlx-1.6.50-py3-none-any.whl — SHA-256 31d2f9ee877220d7147e740f9235275572cf01f20dc69f4aa65c2d8f5172921e
  • Source: vmlx-1.6.50.tar.gz — SHA-256 0cc6ad12b9a60086d941b70c2eaa3ab6d2f7dcc6d9982ef6b095a530db448761