Magic Master 3.10: New Interface, Honest Loudness, Stems

Version 3.10 is about understanding what happens to your track: before
processing, during it and after. The headline is a new interface in public beta,
plus a handful of changes that came straight out of real feedback.

A new interface, in public beta

The new interface lives at /app/beta; the
classic one stays where it is. The path is shorter: upload a track, choose what
to do with it, get the result. Inside:

• AI advice right after upload. The assistant measures loudness, dynamics and
  peaks and tells you which preset fits and why, with no extra button.
• An honest before/after comparison. One button switches between the source
  and the master from the same spot in the track, with loudness matched: louder
  always seems better, so you compare the sound, not the level.
• Result rating. Like or dislike with reasons; every rating reaches us and
  gets read.
• Output format. WAV, WAV 24, FLAC, MP3, OPUS, AAC; 44.1, 48 or 96 kHz; dither
  for 16-bit WAV; bitrate for MP3 and OPUS.
• My masters. Your previous results, downloadable while the file is kept.
• All 24 presets, a custom loudness target, album mastering with shared
  loudness, splitting a track into stems and mastering stems you already have.
• Top up without leaving the page. The track you were about to master waits
  for you after sign-in and after payment.

Beta means we are collecting opinions. There is a "Suggest an improvement" link at
the bottom of every beta page.

Loudness is explained before processing, not after

Half of the negative ratings we went through in early September were not about
sound quality but about unexpected loudness: "too quiet" at a −16 target, "too
bright" after an 8 dB lift. The target LUFS was right there in the field, but
what it meant for that particular track only became clear after mastering. Now
the classic interface:

• shows the cost of the target before you start: the track will get quieter,
  we will lift it, or the punch will soften;
• explains "it got quieter": when the master brought a loud source down to a
  platform norm, a "Need it louder — set −9 LUFS" button is right there;
• explains a loudness miss in words when the safety stage kicked in: the chain
  refused to squash the dynamics any further so as not to ruin the sound;
• warns about AI Unlimiter in advance: on input with a crest factor above
  14 dB it narrowed the dynamics in all 20 real jobs out of 20, and you are now
  told so before processing. The choice stays yours.

Sound

True peak ceiling of exactly −1.0 dBTP. Every page of the service promises
−1.0, while the safety stage at the end of the chain held −0.9: in one week 4.1%
of masters landed above the promise. There is now one ceiling for the whole chain,
and it matches what the pages say.

AI Unlimiter no longer takes back the dynamics it restores. The preset opens
up over-compressed tracks, but the multiband stage after it took back more than it
gave. On seven real files from complaints, skipping that stage was better in six,
and the promise "output dynamics not below input" held in five cases against zero.
On dynamic material the gain reaches +4.2 LU.

Uploads and storage

• Uploads up to 1000 MB for accounts with tokens; up to 160 MB for guests and
  free accounts.
• 30 days of storage for everyone who has bought tokens. The month used to go
  only to the Studio plan.
• A long track is flagged before upload, not rejected after it.
• A file with the wrong extension is accepted. Phones and messengers often
  save M4A as .mp3; in three weeks 38 perfectly good files like that were rejected
  as "unsupported format". The format is now detected from the content.

For guests and newcomers

• All genres are open to guests for free masters. The AI used to recommend a
  locked preset to guests, and one in five people who uploaded a file never got to
  processing.
• At the limit, guests are offered a free account instead of a purchase: an
  account gives you a master every day, thirty a month instead of three, no card.
• AI advice without the wait in the classic interface. When the model was slow,
  people waited up to 45 seconds. Now rule-based advice arrives instantly and the
  model refines it in the background.

Small things you will notice

• When a master is ready, the browser tab title shows "Ready", so you can safely
  switch to another tab.
• Sign-up, address confirmation and password reset emails arrive in the language
  of the page you came from.
• A standalone BPM counter: tap
  the tempo right in the browser, no file needed.
• For people who work through an AI assistant: the MCP server can now master an
  album in one call, preview several presets on 24 seconds of the track, and tell
  you the cost before mastering. If the assistant sends a link to a web page
  instead of an audio file, it gets a clear hint about which link is needed.
  Setup is on the /mcp page.

What's next

What matters most right now is your opinion of the new interface. Try
the beta on your own track and tell us
what works, what is unclear and what is missing compared to the classic interface.