vMLX 1.6.28

Added

• Support for the new Qwen3.6-27B bundle line (and the upcoming Qwen 3.8
  line), whose metadata-only stamps previously killed the server at startup
  or load. The qwen3_coder tool-parser name now resolves to the XML-function
  parser its templates actually emit, capability-only stamps route through the
  stock loader instead of the JANG codec, and the bundled multi-token
  prediction head is constructed, correctly quantized per the stamp's
  per-module overrides, and engaged automatically at the stamp's trained
  speculative depth. Measured on the 4-bit bundle: 23.3 to 31 tokens/sec
  (+33%) with no flags, greedy output byte-identical with the head on or off,
  and image and video probes answering correctly through both API and app.

Fixed

• A prefix-cache hit could change the answer on seven measured model families,
  and the opt-in switch that trades cache reuse for answer stability only
  reached three of them. It was reachable only from inside the family policy
  chain, and minimax, muse_glimmer and step3p7 hold no branch there. The
  switch now runs ahead of that chain, and nanbeige — the family with the
  highest measured divergence rate — has been added to the list it was missing
  from. Proven end-to-end: with the switch armed, both turns reproduce the cold
  answer byte for byte. Absence from the list is not exemption; no family has
  been shown exempt, and the accompanying notes now say so.

• A long-running request could be killed as unresponsive while it was
  generating normally. The engine's liveness counter double-counted output
  tokens, and a recovery restart reset half of the sum, so the value went
  backwards mid-request; the streaming timeout only credits a counter that
  increases. Both the text and multimodal schedulers had the defect. Operator
  logs that report "still progressing (N tokens)" now show the true token
  count rather than twice it.

• Explicitly disabling a parser was silently ignored. --reasoning-parser none
  had no effect on any streaming surface and --tool-call-parser none was
  inert in the module entry point, while the startup banner reported both as
  disabled.

• A derivation step beginning with = rendered as literal dollar-sign text in
  chat instead of as maths.

• The Hugging Face token was normalised on only two of the five paths that read
  it, so a token with surrounding whitespace worked in some places and failed
  in others.

• Muse Glimmer failed to load on a fresh pip install vmlx. mlx-vlm 0.6 began
  shipping its own muse_glimmer package, and vMLX handed the namespace to it
  on sight — but upstream names its preprocessing module differently, so the
  server exited with a missing-module error before it could start, and had it
  started it would have used a forward pass without this port's corrections.
  vMLX now keeps its own validated runtime unless an upstream package genuinely
  provides the same interface. The macOS app was never affected: it pins the
  older mlx-vlm that has no upstream package, which is why the app tested clean
  while the published wheel did not.

• Muse Glimmer failed to load on a fresh pip install vmlx. mlx-vlm 0.6 began
  shipping its own muse_glimmer package, and vMLX handed the namespace to it
  on sight — but upstream names its preprocessing module differently, so the
  server exited with a missing-module error before it could start, and had it
  started it would have used a forward pass without this port's corrections.
  vMLX now keeps its own validated runtime unless an upstream package genuinely
  provides the same interface. The macOS app was never affected: it pins the
  older mlx-vlm that has no upstream package, which is why the app tested clean
  while the published wheel did not.