v0.4.19

What's Changed

Added

• gmlx --help ends with the address of the docs site.
• A docs page shared as a link previews with its title, its opening
  sentences and the gmlx card.
• Qwen3.8-Flash-Next drafts with the llama.cpp mtp-*.gguf MTP files, such
  as those in the MTP folder of the unsloth repo.
• The loader looks for a companion drafter in an MTP folder inside or
  beside the model's folder, as well as beside the model.
• /v1/systemone and gmlx systemone answer on OpenJev and any other text
  model with OpenJev's letter readout, and POST /v1/prewarm reads a
  decision state ahead of its questions.

Changed

• The prompt cache stores and restores checkpoints of models with
  recurrent layers faster.
• /v1/systemone reports the official Jev confidence for choice and score
  answers in place of the chosen probability. think_threshold still
  compares the chosen probability, which diagnostics.think_auto now names
  chosen_probability.
• The README links each page to the docs site instead of the Markdown file
  on GitHub, and the Homebrew formula names the site as its homepage.
• gmlx run and gmlx chat turn speculation on by themselves when they
  find the companion drafter of a Qwen3.8-Flash-Next or Muse Glimmer model,
  as they already did for DeepSeek-V4.
• Requires mlx-kquant 0.4.15. With it, gemma-4 31B and 26B-A4B decode
  faster, and so does plain decode of Q6_K models under sustained load.
• Gemma-4 decodes faster on NAX GPUs (M5) beyond a few thousand tokens of
  context. GMLX_GQA_SDPA_NAX512=0 opts out.

Fixed

• --kv-bits keeps an fp16 cache on Kimi K3, whose attention cannot read a
  quantized one. The server quantized its attention layers and failed at the
  first token, and run and chat did the same on Kimi K2 and K3.
• GLM-5.3 runs with --kv-bits. A quantized cache failed at the first
  token in its attention layers.
• gmlx launch sends the config's server.api_key to a keyed server that
  is already running. Before, only a server that launch started got the
  key, and a running one refused the launch.
• gmlx validate reports a Hadamard-folded file as not loadable when the
  loader cannot run its fold version or architecture.
• gmlx talk refuses a speech speed outside 0.25 to 4 from --speed,
  /speed or talk.speed. The server refused to speak at such a speed.
• /thinking-budget in gmlx chat no longer warns that MTP replies ignore
  it, since they honor it.
• gmlx init no longer offers to install the [embeddings] extra, which
  installs nothing.
• Help text: /help in gmlx chat lists /thinking adaptive, serve --tts lists the qwen3-tts-small alias, and gmlx talk names the
  command to run when the server has no default model.

Full Changelog: v0.4.18...v0.4.19