Skip to content

vMLX 1.6.57

Latest

Choose a tag to compare

@jjang-ai jjang-ai released this 11 Sep 03:26
· 1 commit to main since this release

vMLX 1.6.57

  • Improved local-folder discovery and launch validation for compatible mflux image generation/editing exports, including quantization metadata, progress, cancellation and output history.
  • Corrected downloader cancellation, paused/queued states and local-storage error handling.
  • Corrected output-token settings, live settings refresh, tool-error finalization and gateway usage reporting.
  • Added Spark-X2.5 integration and corrected mixed-state and Flash-Next media-prefix continuation handling.
  • Corrected terminal reasoning/content separation and rejection of malformed Gemma native tool envelopes without fabricating tool arguments.
  • Native MTP recovery and demotion use measured execution cost. The selected fixed depth is a ceiling; an instantaneous or universal autoregressive speed floor is not guaranteed.

Model-generated answers can still be incorrect. Media output can vary between cached and uncached execution; universal numerical equivalence and all-family quality are not claimed. Experimental kernel optimizations remain opt-in. Additional GLM/MTP performance and broader quantization qualification continue separately.

Known limitation: a connected Gemma video/API test omitted a requested tool call and produced an incorrect answer; its cause remains under investigation. Successful transport is not a guarantee of model answer accuracy. The intermittent 27B slowdown and model-dependent media re-prefill remain follow-up work.