MiniMax-Music3 vs Lyria 3.5: Open Weights Against Google's Walled Music Garden

MiniMax-Music3 and Lyria 3.5 both generate complete songs from a text prompt, and that is roughly where the similarity ends. G​ogle dropped Lyria 3.5 into its Flow Music consumer app on July 29, 2026, free to subscribers, with no weights and no confirmed developer API; Mini​Max released MiniMax-Music3 as open weights on Hugging Face on August 13, 2026, after its music-3.0 API had been live since July 16. Same job, opposite distribution strategies. This comparison is about what that difference actually buys you, and it starts with a sourcing caveat: every quality claim on both sides is vendor-reported, and neither model has an independent benchmark or a published head-to-head as of writing.

Two release philosophies

Lyria 3.5 is a consumer feature. It lives inside Flow Music (Google's acquired-and-rebranded Riffusion product), generates tracks from 30 seconds up to three minutes, and is described by Google as delivering "more realistic and emotionally nuanced vocals." It supports multi-language vocals — the launch demos span English, Spanish, French, Hindi, and Portuguese — plus image-to-music input, controls for tempo, BPM, key, vocal style, and instrumentation, and a "Selective Section Painting" tool for reworking part of a track. Google says the model page carries a SynthID watermark.

MiniMax-Music3 is a developer artifact. The open-weights release is a full pipeline — 8B global LLM plus 0.6B local LLM fused through a 2.4B Flow Matching module into a Flow-VAE that outputs 32 kHz stereo WAV — and it is self-hostable in roughly 22 GB of VRAM. Its control surface is a structured caption: genre, subgenre, BPM, key, scale, emotional progression, vocal details, and a per-section arrangement map, with lyrics tagged [Verse], [Chorus], [Bridge], and so on. If Lyria 3.5 is a product, MiniMax-Music3 is a component.

The scoreboard

• Song length — MiniMax-Music3 up to five minutes in a single generation vs Lyria 3.5 up to three minutes.

• Weights — MiniMax-Music3 open, downloadable since 2026-08-13, no gate vs Lyria 3.5 closed, no weights, no self-hosting.

• Output — MiniMax-Music3 32 kHz 16-bit stereo WAV files you keep and own vs Lyria 3.5 audio that lives in Flow Music's player.

• Control — MiniMax-Music3 structured caption covering BPM, key, scale, arrangement per section vs Lyria 3.5 tempo, BPM, key, vocal style controls plus Selective Section Painting after generation.

• Serve it yourself — MiniMax-Music3 yes, documented at ~22 GB VRAM with SGLang-Omni, diffusers, and ComfyUI paths vs Lyria 3.5 no.

• Price — MiniMax-Music3 $0.15 per song (up to five minutes) on the API, free tier at 3 requests per minute vs Lyria 3.5 free inside Flow Music subscriptions, with developer API pricing still not confirmed for 3.5.

That last line needs a footnote. The ~$0.04 per 30-second clip and ~$0.08 per three-minute song figures circulating for Lyria are reported for the earlier Lyria 3 and 3 Pro generations, not confirmed for 3.5, and Google has not announced a model ID or endpoint for 3.5 on the Gemini API. Treat Lyria's API pricing as an open question.

The MiniMax-Music3 model card shown above is the whole documentation for the open release — there is no separate paper or launch post to point to, which is why this comparison leans so heavily on the README's claims.

The vocals question

Both models lead with their vocals. Google claims "more realistic and emotionally nuanced vocals, plus improved pronunciation" for Lyria 3.5; MiniMax's release notes for music-3.0 claim a new vocal engine that removes high-frequency "digital hiss" and controls melody, articulation, breathing, and layered harmonies. Both are vendor claims, and neither has been reproduced independently.

There is, however, one early body of independent testing for Lyria 3.5, and it is genuinely mixed. A community test on Linux.do found vocals and melody improved but Chinese lyric recognition regressed to roughly 5% accuracy; a Japanese-language hands-on found convincing vocals with occasional pronunciation quirks. That is the honest baseline for Google's pronunciation claims: convincing in some languages, shaky in others. For MiniMax-Music3 there is no independent test at all yet — the repo has been public for a day — so the only fair statement is that both sides' vocal claims are unverified, and Lyria 3.5's has already been partly contradicted. We covered the Lyria side of this in more depth in our earlier Lyria 3.5 launch write-up, shown below.

What each one actually costs

The cost comparison is lopsided depending on how you count. As a consumer, Lyria 3.5 is effectively free — Flow Music's entry tiers are around $6 to $8 per month and the model is included. As a developer, MiniMax-Music3 is dramatically cheaper: $0.15 per five-minute song, or the cost of a GPU if you self-host, versus Lyria's unconfirmed API pricing and the fact that the API may not exist yet.

The per-minute math is worth doing explicitly. MiniMax's $0.15 buys up to five minutes of song, so roughly $0.03 per minute at full length. Even the most optimistic reported Lyria figure, ~$0.08 per three-minute track, is about $0.027 per minute — nearly identical. The real difference is not the unit price; it is that MiniMax's price is published and the model is runnable on your own hardware, while Lyria's developer story is speculative.

Who should pick which

Choose Lyria 3.5 if you are a Flow Music subscriber, you want the polished consumer experience, you value Google's ecosystem (image-to-music, editing tools), and you are comfortable with output staying inside the app and with no weights and no confirmed API. It is the better product for people who want to make songs, not operate a model.

Choose MiniMax-Music3 if you are building something — an app, a game soundtrack pipeline, a cover generator, a lyrics-to-song service — where the output has to be a file you control, the cost has to be predictable, or the model has to run where you run it. The open weights and the structured caption give it a decisive engineering advantage for integration work, and the five-minute ceiling is a real edge over Lyria's three.

For teams that want neither bet locked in, the pricing hedge matters as much as the models: on a pass-through router like OrcaRouter, provider list price is passed through with no markup, so when a vendor cuts price the cut is live the same day — no renegotiation and no repricing from a middleman. That is worth remembering in a category where both vendors' pricing is this young.

The honest bottom line

Lyria 3.5 is the better consumer experience with the bigger brand and the unproven API; MiniMax-Music3 is the better developer asset with the published price, the open weights, and zero independent listening tests. For the next few weeks, that asymmetry is the whole story: Google is selling a feature, and MiniMax is shipping a foundation. The question that will decide this matchup is whether the open checkpoint sounds as good as the API version — and nobody has published that answer yet.