Skip to content

vMLX 1.6.56 — Console Amber

Latest
Latest

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 08 Sep 20:01
· 1 commit to main since this release
v1.6.56
c006a36

vMLX 1.6.56

Console Amber

  • New Chat & Images, Servers & API, and Models navigation, with the updated app icon, readable amber contrast, responsive controls, and General Preferences.
  • Existing chats, sessions, and saved settings are preserved. New profiles default to one loaded model; existing profiles retain their choice.
  • Server settings group concurrent processing, SSD prefix caching, power, generation, MCP, and speculative decoding. Removing retired controls does not silently reset existing settings.
  • JANG model feeds filter out non-JANG repositories. Update notices remain usable in narrow windows and a dismissed version does not hide later updates.

Runtime and API

  • Atomic on-disk safetensors alignment repair during local model preflight, including nested draft/media shards, with validation and unsafe-path protections.
  • MiniCPM5 native tool/reasoning integration and guarded affine projection handling.
  • Per-request image/video controls, effective-setting diagnostics, and typed strict-mode errors across API dialects.
  • Hybrid partial-prefix and media-position cache corrections, idle SSD cleanup, capacity-only notices, and guarded cache clearing.
  • Tool/JSON parsing preserves string data and terminal punctuation; Ollama generation streams retain final usage without repeating content.
  • MTP context accounting includes restored tokens. Existing depth policies remain in effect.

Limits

  • No universal MTP-over-AR speed floor is promised. The previously observed intermittent Qwen 27B 2D slowdown remains unresolved.
  • Media controls can reduce sampled content, and cache reuse depends on exact prompt, media, and settings identity. New media may require prefill.
  • Schema validation cannot guarantee that a model generates a requested tool call or repeats every requested literal. Warning transport differs by API dialect.
  • The packaged MiniCPM check covered connected UI and Chat/Responses tool continuations. Its Responses final answer omitted a requested repeated filename; tool arguments, results, streaming, and cache ordering were intact. This is not an all-model quality or compatibility guarantee.

Downloads and provenance

Both macOS DMGs are Developer ID signed, independently notarized, stapled, and Gatekeeper checked. Tahoe is the macOS 26 build; Sequoia is the compatibility build.

  • Source: c006a367e8a72e6fb1f4c158f1272cfccf1ea2ee
  • Bundled JANG: 2.5.47 at f3c79081c3d99e8a26fa28444934f7567acb00c6
  • Tahoe SHA-256: 4c5ea8634443daee2f4ed143c20609eb1304c75b501e9f433ff3e204fb2728dc
  • Sequoia SHA-256: c14c83bc70bcf378e01a4cf33305c6be1f3bc72d852f1e0fe823e2ff4a07cf99

Wheel and source distribution are included as release assets. PyPI publication is tracked separately.

Assets 9
Loading