Skip to content

vMLX 1.6.10

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 15 Jul 07:25
· 1885 commits to main since this release
v1.6.10
8406238

vMLX 1.6.10 ships the urgent Bonsai affine-1 runtime bridge and paged-cache L2 reliability fixes.

Highlights:

  • Loads schema-2 JANG affine 1-bit and ternary text/VLM bundles with per-module quantization overrides.
  • Preserves Qwen 3.5 hybrid GDN/SSM + attention cache routing, VL/video tensors, TurboQuant attention KV, and async SSM rederive.
  • Enables paged block-disk L2 cache by default with an explicit opt-out and matching Electron UI defaults.
  • Fixes continuous-batching request cleanup so free-block eviction and L2 writes occur under pressure.
  • Keeps local Electron chat on the OpenAI Responses API by default and derives chat sampler defaults from model bundle configuration.

Verification:

  • Source commit: 693b2d0.
  • Engine focused release suite: 276 passed.
  • Electron panel suite: 2,206 passed, 3 skipped; TypeScript typecheck passed.
  • Final signed/notarized Sequoia app loaded Bonsai-27b-1bit-JANG and completed three-turn Auto/Off/On chat over /v1/responses with visible reasoning separation and paged+SSM cache reuse.
  • Native run_command invocation executed in the final app; post-tool continuation remains a known model-specific follow-up because Bonsai repeated an empty continuation and the run was stopped.

macOS DMGs are published in the companion jjang-ai/mlxstudio v1.6.10 release.

Assets 2
Loading