Skip to content

vMLX 1.5.58 Checkpoint

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 12 Jun 00:30
· 34 commits to main since this release
v1.5.58
08d8fcc

vMLX 1.5.58 Checkpoint

This is a signed and notarized checkpoint release. It is intended to make the
current vMLX Python engine and MLXStudio app fixes available while several
larger model-family rows remain under active proof.

Highlights:

  • Refreshes the bundled vMLX Python engine to 1.5.58.
  • Bundles the current JANG runtime package surface from jang==2.5.31.
  • Keeps current parser/API fixes for Qwen/Qwen-coder Responses tool-call
    streaming, missing required argument fail-closed behavior, function-call
    argument deltas/done events, output-index validation, and tool-result
    continuation.
  • Keeps the current Gemma4/JANG/QAT/MXFP runtime and bundled app parity fixes,
    including Gemma4 unified runtime registration and JANG/JANGTQ loader parity.
  • Keeps current cache/runtime packaging parity for JANG/JANGTQ/TurboQuant
    kernel imports and block-disk/cache surfaces included in the bundled engine.

Verification:

  • panel/scripts/verify-bundled-python.sh passed before packaging.
  • panel/scripts/build-release-dmgs.sh all produced fresh 1.5.58 Sequoia and
    Tahoe DMGs under explicit checkpoint override.
  • Apple notarization accepted both DMGs.
  • panel/scripts/verify-release-dmgs.sh passed for both DMGs: disk image
    verification, Developer ID signature validation, stapled ticket validation,
    and Gatekeeper assessment all passed.

Artifacts:

  • vMLX-1.5.58-sequoia-arm64.dmg
    • SHA256: 71925fa21857a631c7fdddfd14b217cef8e076a3ce88fb82439672a0196bd7f4
    • Notary ID: d94933ed-ba35-4d45-a4f5-5e9c8ce8ff1a
  • vMLX-1.5.58-tahoe-arm64.dmg
    • SHA256: ffa671547b0de037d9e5257589f29d8e29c5cebb7358c127ed0a90b6925040dc
    • Notary ID: 8615b942-0b30-4aea-a805-f62c49895ca2

Known checkpoint boundary:

  • This is not a full production-clear model matrix release.
  • Open rows remain in the checkpoint override manifest for larger live
    model-family proof, selected UI/media rows, and deployed/tunnel parity
    surfaces.
  • PyPI publishing is separate and remains gated on valid PyPI publishing
    credentials or trusted-publishing configuration.

Credit: thanks to @hornsan1 for reported runtime/API issues that informed this
checkpoint hardening pass.

Contributors

hornsan1
Assets 6
Loading