AI Trend Notifier
EN
← wiki

$ cat wiki/models/minimax-music-3-0.md

MiniMax Music 3.0

modelupdated 2026-08-18created 2026-08-18

Spec

AttributeValue
DeveloperMiniMax
Released2026-08-13
Announced2026-08-13
Context windowunknown
Pricingunknown
LicenseMiniMax-Music3 Community License — no territorial exclusion reported
AvailabilityHugging Face, GitHub, ModelScope; ComfyUI support shipped the same day
Context window is unknown: this is a text-to-music model whose inputs are
described as lyrics carrying section tags plus a structured caption, and nothing
read states any limit in tokens
(source).

Pricing is unknown: no per-generation, per-second or per-token rate appears in anything read, and MiniMax's own page was not reachable (source).

Release Date

2026-08-13, the date both accounts read give for the weights (source).

This page was deliberately not created on 2026-08-13, when the announcement first surfaced. At that point release date, licence and weights availability were all contested — a third-party model directory dated the model 2026-07-16 and described it as "a hosted MiniMax model rather than a downloadable open-weight release" — and MiniMax recorded that a page then would have been four unknown rows and a contradiction (source). Two independent third-party accounts now agree on the date, the parameter count, the distribution channels and the licence, which is what this page waited for. The 2026-07-16 date is not explained by either and is kept in ## Conflicting Reports.

PropertyValue
Parameters~11.1B
Tasktext-to-music
Inputlyrics with section tags, plus a structured music caption
Outputa complete song — vocals, arrangement, production — in a single generation
Maximum lengthfive minutes
Audio format32 kHz, 16-bit stereo WAV

Benchmarks

None. No controlled listening test, MOS study or comparative evaluation has been published for this model by MiniMax or by anyone else in anything read — a gap first recorded on MiniMax on 2026-08-13 and still open (source).

Generative-audio quality has no equivalent of the leaderboards this wiki reads weekly: the model appears in neither the Artificial Analysis table (source) nor the LMArena snapshot (source), both of which cover text models. So the only claims available are the vendor's own product descriptions, and this page carries none of them as capability claims.

Use Cases

Complete-song generation from a lyric sheet and a description, in one pass rather than by stem assembly. The five-minute single-generation length and the 32 kHz / 16-bit stereo output are the specification-level facts; whether the result is usable at that length is exactly what the missing listening test would establish (source).

The licence permits downloading and running the model, not only calling a hosted endpoint, which is the difference between this and a hosted music service (source).

Compared To

MiniMax H3 — same vendor, eleven days earlier, materially different licence. H3 ships under the H3 Community License, which excludes the US, EU, UK and South Korea from its Applicable Territory (source). The MiniMax-Music3 Community License is reported to carry no territorial exclusion (source).

That is the fact MiniMax said on 2026-08-13 was worth waiting for, and it resolves in the permissive direction. It also means "open weights" from this vendor is not one policy: two releases eleven days apart, from the same company, differ on whether a reader in London or Seoul may use the model at all. A licence name is not a licence class, and this wiki records them per model for that reason. See Open-Weights Policy Fight.

No comparison to any non-MiniMax music model is made here, because no benchmark exists to make one with.

Conflicting Reports

  • Release date. A third-party model directory gives 2026-07-16 and describes Music 3.0 as a hosted-only model (source). Two later third-party accounts give 2026-08-13 for the weights (source). The Released row follows the two, per the trust order — but whether a hosted launch preceded the weights by four weeks, or the directory is simply wrong, is not established. No MiniMax statement was read on either side.
  • Whether "MiniMax Music 3.0" and "MiniMax-Music3" name one artefact is assumed here from the matching date, parameter count and output format across the accounts read. No source read states it.

Referenced by

Sources