Skip to content

vMLX 1.6.12

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 20 Jul 06:30
· 1696 commits to main since this release
v1.6.12
255821a

vMLX 1.6.12

Cache-hierarchy, streaming-recovery, and gateway-lifecycle checkpoint.

Highlights

  • Makes block-disk L2 a usable prefix tier even when paged RAM is disabled,
    while preserving the RAM-first then disk-refault hierarchy when both tiers
    are enabled. Partial-prefix restore, eviction, immediate-stop durability,
    and restart refault are covered by the current scoped evidence rows.
  • Hardens q4 TurboQuant storage for eligible KV components and hybrid
    attention paths without conflating it with model-weight formats. JANG
    affine, JANGTQ/MXTQ Hadamard-codebook, and base MLX MXFP remain distinct;
    DSV4, openPangu, MiniMax-M3, Gemma mixed-SWA, and other typed/native cache
    paths retain their architecture-specific policies.
  • Fixes abandoned non-stream gateway ownership across Chat Completions,
    Anthropic Messages, and Ollama so a disconnected client cancels the upstream
    request before headers and an immediate follow-up can recover normally.
  • Tightens progressive reasoning/content/tool streaming and terminal/usage
    ordering across current Qwen, HY3, Step, MiniMax, Anthropic, Ollama, and
    Responses routes, including no-tool prompts and rejected/incomplete native
    tool markup.
  • Improves model-derived UI truth for quant labels, parser/reasoning/cache
    settings, single-model swaps, session PID/sleep lifecycle, eager Start
    materialization, port/LAN rollback, and missing-model repoint/removal.
  • Extends current media ownership and cache-key coverage for Step and Nemotron
    Omni image/video/audio paths while preserving MiniMax-M2.7 as text-only.

Verification

  • Runtime source checkpoint: 6de1096eca0ea2d5516ad64d6e79da98f3ae20a2;
    tag: v1.6.12.
  • Full Python suite: 6,186 passed, 185 skipped.
  • Full panel suite: 2,332 passed, 3 skipped; TypeScript typecheck passed.
  • Both Sequoia and Tahoe DMGs passed bundled-runtime verification, Developer ID
    signing, fresh Apple notarization, stapling, Gatekeeper verification, and
    exact installed-app smoke.
  • The Sequoia app completed a three-turn UI reasoning/tool/recall loop, raw
    Responses and Chat streaming, and a live
    paged+mixed_swa+tq-native cache hit. Tahoe independently completed a UI
    reasoning/content turn and a raw Responses stream with separate reasoning
    and content deltas, terminal completion, and usage.

macOS builds

  • Sequoia/Sonoma-compatible build (macosx_14_0_arm64; macOS 14.5+):
    vMLX-1.6.12-sequoia-arm64.dmg
  • Tahoe-native build (macosx_26_0_arm64):
    vMLX-1.6.12-tahoe-arm64.dmg

Checksums

  • Sequoia SHA-256:
    704d87edf168a73d4ca2d94e8cb6190ca593ada71bca181bf369c84ea13ae421
  • Tahoe SHA-256:
    81b9205a722282cc1eec75713c18dec3efc34ed76e3bcaf6587147e0ce372c49

Known follow-up

This is a tested release checkpoint, not closure of every retained family or
stress row. Broader signed-app model repetition, longer stochastic and media
soaks, remaining parser-family tool rows, injected mid-stream failure, and the
documented Laguna/Bonsai/Step/DSV4/M3/openPangu/LFM/Nemotron follow-ups remain
on the post-release worklist.

Assets 2
Loading