Skip to content

vMLX 1.5.63

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 18 Jun 10:11
· 2055 commits to main since this release
v1.5.63
b574e5c

vMLX 1.5.63

Highlights:

  • MiniMax-M3 compatibility update focused on REAP40-d3 loading, multimodal routing, reasoning controls, tool calls, and cache-hit stability.
  • Gemma 4 E2B/E4B/12B MXFP4 vision/text compatibility update with live UI and API coverage.
  • Model-owned generation defaults are preserved for supported startup paths.
  • MiniMax-M3 keeps its native MSA/Lightning sparse cache policy: paged cache off, TurboQuant KV skipped for native MSA, and JIT disabled for the dynamic sparse-cache path.
  • Gemma audio is not advertised in this release unless explicitly proven by the artifact/runtime.

Verification:

  • vMLX source commit: 12382f2.
  • MM3 installed-app stress covered 10 UI turns, reasoning off/on/auto, tool calls, image input, long-context recall, prefix-cache hits, Chat/Responses/Anthropic/Ollama, and streaming tool/image surfaces.
  • Gemma 4 E2B/E4B/12B MXFP4 installed-app stress covered 10 UI turns, reasoning off/on/auto, image input, tool calls, cache hits, Chat/Responses/Anthropic/Ollama, and streaming rows.
  • Sequoia and Tahoe DMGs are Developer ID signed, Apple notarized, stapled, and Gatekeeper accepted.

Artifacts:

  • Sequoia DMG SHA256: 0be2ad302ae391efb7a1af2203db831162f60ec8271bd30a071a8c48e303c0dc
  • Tahoe DMG SHA256: aff1cc423dff786282d7f8e9fc2e1022d763b435a4e2d6e09bc941394d768a30

Scope:

  • This is a scoped MM3 + Gemma 4 VL compatibility release. Broader model-family matrix work continues in the next release line.
Assets 6
Loading