rapid-mac-v0.12.5

[0.12.5] — 2026-08-05

A stability release. The headline is a crash: a reply with a formula in it
took the whole app down, and an ordinary maths question was enough to
produce one. The starter model that shipped with 0.12.1 also turned out to
be unusable, so it has been replaced.

Fixed — crashes and hangs

• Maths no longer crashes the app. When a reply contained a formula, the
  app quit outright with a macOS crash dialog — a plain 2 + 2 = 4 was
  enough. Formulas now show as their plain-text source instead of taking the
  app down. (Typesetting them properly is still to come.)
• A silent server no longer hangs the chat. Sending a message used to sit
  there indefinitely if the model stopped responding; it now fails after a
  bounded wait and the message stays retryable.
• Quitting no longer stalls or crashes, and starting the app no longer
  risks shutting down a Rapid-MLX server you started yourself in a terminal.

Fixed — first run

• The starter model has changed. The previous one fell apart on ordinary
  multi-step questions — doubling words, then looping until it ran out of
  room. Measured on the same Mac and the same question: 0/4 correct before,
  16/16 after, with the first answer arriving in about a second.
• Two models that could fail to answer are temporarily hidden. Ministral 3
  and Gemma 4 E2B looked like ordinary small chat models in the picker, but on
  some Macs sending them a message produced no reply at all, or a reply that
  made no sense. They stay hidden until that is fixed.
• 8 GB Macs get a recommendation instead of a rejection, and every RAM
  size now gets a "smart" and (where it helps) a "fast" pick.

Fixed — chat and models

• Conversation history keeps its order. Older chats could jump around in
  the sidebar.
• The compose box grows with what you type, and the transcript stops
  yanking you back to the bottom while you are reading further up.
• Loading a model that will not fit is now blocked with a warning based on
  free memory at that moment, rather than freezing or panicking the Mac.
• "Speed on this Mac" shows live progress and an estimate instead of an
  unmoving spinner, and the size column no longer mixes download size with
  memory use.

Fixed — connecting your tools

• Your API key is no longer shown in the clear in the copyable setup
  snippets on the Launch page.
• Cursor no longer gets a configuration that cannot work. Cursor routes
  requests through its own servers, so a localhost address was never
  reachable; the app now says so and points at Claude Code, Cline, or
  Continue for a local connection.

Changed

• A refreshed visual system for the model list and its states.
• Bundles the Rapid-MLX 0.12.4 engine. What you should notice: replies
  are faster on a long conversation because the model no longer re-reads
  what it has already read; models that think before answering keep their
  reasoning out of the answer; the app can tell whether a model is busy or
  idle; and several models that used to fail to start — including the
  Gemma 4 family — now load. LFM2.5 models are newly available.