Cognitum Music turns lyrics and a production brief into a full arrangement — intro, verse, chorus, bridge, outro — up to five minutes long, with a consistent vocal identity from the first bar to the last. Built on MiniMax-Music3, self-hosted and served through an authenticated Cognitum API on GCP Cloud Run GPU. Twelve tracks below, produced by the model this page describes.
Nothing below is templated — each track is a distinct lyric sheet and structured production brief, sent to the same weights. Scroll to hear how the brief shapes the output.
Eight more tracks. Click any card to hear it — each mini-waveform is the real generated audio, not a placeholder. Includes a 15-second ad jingle and a classical chamber piece with no vocals at all.
Everything above is a 15–20 second excerpt. This is the real deliverable — intro, two verses, three choruses, a bridge, and an outro, generated as a single continuous pass with the same vocal identity and an arrangement that actually builds across it, not eight clips glued together after the fact.
1:40, 8 sections, one uninterrupted generation. BPM 100, E minor rising to E major at the chorus.
Initialized from Qwen3-8B, adapted to semantic music tokens. Predicts one frame at a time, holding the song's long-range structure — verse, chorus, bridge — coherent across five minutes.
Fills in the remaining seven residual RVQ codebooks per frame — the fine-grained acoustic detail the Global LLM's single semantic codebook doesn't carry.
Fuses both models' hidden states directly — continuous, not discrete tokens — preserving vocal articulation and timbre that a token-only decode would lose.
Adapted from MiniMax Speech, retrained for music's dynamic range. Outputs 44.1kHz stereo audio directly from the fused latent.
This is a batch model — one pass produces a complete track. There is no
streaming generation, no live parameter steering. Cognitum
Music's /v1/music/stream endpoint streams the
delivery of an already-finished file over chunked HTTP — first
bytes sooner, not faster generation. Full precision needs 24GB+ VRAM;
every sample on this page was produced locally on a 16GB consumer card
via group-offloading, which is why the loudness and stereo width lean
conservative — the production Cognitum Music API runs full precision
on a dedicated Cloud Run GPU instance instead.
The gateway validates a Cognitum cog_-prefixed bearer key — the
same scheme already used across Cognitum's agent services — meters requests
per account, and proxies to an internal-only GPU inference service.
Blocks until the complete WAV is ready, then returns it in one response. Generation is minutes, not seconds — size client timeouts accordingly.
curl -X POST https://minimax-music-gateway-63rzcdswba-uc.a.run.app/v1/music/generate \ -H "Authorization: Bearer cog_..." \ -H "Content-Type: application/json" \ -d '{ "lyrics": "[verse]\nMorning light...", "prompt": "Genre: acoustic pop. BPM: 96...", "audio_duration": 60.0, "seed": 7 }' --output song.wav
Same request shape. Response arrives as chunked HTTP — the client can start reading before the full file is buffered. Progressive delivery, not real-time generation — see the constraint note above.
# identical body to /generate — # only the response transfer differs curl -X POST https://minimax-music-gateway-63rzcdswba-uc.a.run.app/v1/music/stream \ -H "Authorization: Bearer cog_..." \ -H "Content-Type: application/json" \ -d @request.json --output song.wav
Returns remaining daily generations for the authenticated account. Metering is per-account, per-day, per-gateway-instance — disclosed as such, not a distributed quota system.
curl https://minimax-music-gateway-63rzcdswba-uc.a.run.app/v1/music/usage \ -H "Authorization: Bearer cog_..." → { "account": "musica-desktop", "generations_remaining_today": 17 }
Each tag goes on its own line inside lyrics — text sharing a line with a tag is dropped by the model's input contract. Nine tags cover every part of a song structure: