Skip to content

Latest commit

 

History

History
35 lines (28 loc) · 1.96 KB

File metadata and controls

35 lines (28 loc) · 1.96 KB

MLX Workshop 0.1.0-beta.2

Source prerelease. The DMG is intentionally withheld until its Developer ID signature, notarization, stapling, and Gatekeeper checks pass. Do not redistribute an ad-hoc local build.

This first beta focuses on one trustworthy path: inspect a compatible local float model, review an exact uniform quantization plan, create a new candidate, and qualify it without changing the parent. A guided five-step flow leads from folder setup and plan review through verification and immutable local release staging.

Supported conversion modes are MXFP4, MXFP8, and affine. Already-quantized models, unknown tensor semantics, and unsupported model families fail closed. Measured Llama mixed precision and eligible Qwen3.5 behavior experiments are live but experimental and unqualified. Vision is an advertised-VLM smoke test, and MTP support is a read-only MTPLX compatibility check; vision grafting, automatic MTP packaging, and canonical model replacement remain outside the beta.

The app is Apple Silicon-only, requires macOS 14+, and includes its complete locked Python/MLX runtime. No separate developer tools are needed for the notarized DMG.

Verification evidence:

  • 67 Python unit/integration tests
  • 101 Swift unit/integration tests
  • 2 native UI journeys covering fresh setup and every live interface section
  • Real tiny-model conversion with source immutability and three hashed qualification gates
  • Real measured tiny-model mixed-precision analysis and reloadable materialization
  • Fully hash-manifested 70-package bundled runtime with import/host smoke tests
  • Real end-to-end conversion through the runtime embedded in the Release app, with a regression gate proving child inference cannot write bytecode into the signed bundle
  • arm64 Release bundle, sandbox/privacy checks, and verified drag-to-Applications DMG

Completion is not qualification. The app keeps those states visually and semantically separate, and never starts or resumes a run on launch.