Show IH: I built an offline AI music generator for iPhone.
Spent the last 6 months, solo, trying to get Magenta RT - a real-time music generation model running entirely on an iPhone. No server, no API, no internet after the first download. Open the app and it generates on-device.
The model itself already existed, built for other people's servers. Getting it to run on a phone at all, using Apple's on-device chip instead of the cloud, turned out to be its own project. Way more than just "export the model and ship it." There were a lot of dead ends where things technically worked but were unusably slow, or worked on my phone and not others, or just didn't run at all and I had to figure out why.
I know it's not going to touch what Suno or Udio can do quality-wise, they've got real infra behind them. But I wanted to see if a phone making music with nothing else involved was even possible to ship, not just a cool idea.
Being honest about where it's at: generation still takes a while and isn't gentle on the battery, so "runs on your phone" doesn't mean "runs fast" yet. The model also has to download on first launch and it's around 1.5gb. And I've only really validated performance on iPhone 15 Pro and 17 Pro so far. Older hardware is still a question mark.
Link: www.murmurapps.com
Curious if anyone else here has fought to get something ML-heavy running fully on-device?
On the revenue: it's a one-time 14.99 EUR purchase, not a subscription, but 2 people who aren't me or friends have paid for it since launch 3 weeks ago. Small, but real.
I like that you were clear about the tradeoffs instead of pretending they don't exist.
That honesty actually highlights an interesting positioning choice. If someone buys this because it's the best AI music generator, you'll lose to the cloud. If they buy it because it's always available, private, and independent of a server, you're competing on something entirely different.
Thanks for the feedback!
That’s exactly the angle I’m aiming for. I want to make sure expectations are clear from the start, which is why the website specifies it’s strictly instrumental and includes direct audio examples.
I also built a generous free version so people can fully test how it runs on their specific hardware before spending a dime. If they decide to buy, it’s a conscious "yes, I need this" decision.
And on the pricing—I believe we need more apps that reject the subscription model. For a utility that runs entirely on your own hardware without ongoing server costs, a one-time purchase just felt like the right call.
Glad it resonated.
Your reply made me think there's one strategic decision sitting underneath that positioning which becomes much more significant as the product grows, but I don't think I can explain the reasoning properly in a thread without oversimplifying it.
If you're interested, what's the best email to reach you on?
Thanks, I would love to hear your input. You can reach me here: [email protected]
Thanks! I’ve just sent it over.
Looking forward to hearing your thoughts whenever you have a chance.
The 1.5GB download + performance constraints feel like the real bottleneck for adoption right now. Are you thinking about smaller distilled models or different tiers based on device capability?
Yeah, I currently run two models: MRT1 Large (the big one) and MRT2 Small.
MRT2 Small is much better for speed and download size, but the quality doesn't really compare to MRT1 Large. On top of those, there are the decoders and the text input reader model.
Because of all that, the footprint quickly adds up to around 1.5 GB—and that's even after quantization.
6 months solo to get something running fully on-device is wild, that's the kind of thing that quietly separates "cool idea" from actually shipped.
the honesty here is what makes it land, calling out the battery, the 1.5gb download, the untested older hardware. way more trustworthy than the usual launch hype.
2 strangers paying is real, that's proof someone wanted it enough without knowing you.
Two paid users in 3 weeks for a local-only AI music generator is a fantastic start! There is a huge market of people who hate subscriptions and value local privacy. Total respect for the engineering hustle it took to get this running on-device instead of just wrapping a standard API.
Thanks a lot! I appreciate the support.
Two strangers paying is a stronger signal than most founders realize; friends lie, strangers pay. The asset you're underusing is the 6-month war story itself: a technical writeup on getting Magenta RT running on-device would outperform any ad you could buy, because the people who care about on-device ML are exactly the people who'd pay for this. Ship the blog post before the next feature.
Thanks, good point! Any thoughts on where's best to post it, here, or somewhere on reddit?
Really impressive work! On-device AI music generation is a tough technical challenge — Core ML model optimization for mobile constraints is no joke. Have you considered expanding to other audio formats like podcasts or sound effects? Also curious about the monetization model — is it one-time purchase or subscription?
Thanks! It's an optional one-time purchase for the Pro version or a generous free tier
The honest tradeoff section is the strongest part of the post.
On-device AI is almost a different category from "cheaper cloud AI." The hard question is not just whether the model runs locally, but which constraint the buyer cares about most: privacy, offline use, latency, battery, or quality.
If you had to pick one wedge for the next version, would you rather make generation faster on newer phones, or make the offline/privacy promise clearer to the users who already value it?
Thanks! I'd lean toward making generation faster. The privacy aspect is already a built in truth of the architecture, but improving performance turns it into a daily utility. Plus, with Apple's hardware advancements every year, the baseline performance is only going to get better on its own.
Very good idea and It is an AI Music Generator
the "worked on my phone, not others" line is the part id worry about most long-term. on-device means apple's entire device matrix becomes your QA surface — every chip gen has a different neural engine, memory ceiling, thermal budget. the cloud version has one runtime you control; you have every iphone since the A-whatever, forever. shipping it was the hard part, but keeping it running across devices you dont own is the tax that never stops. impressive that you got it running at all though.
Yeah, the QA part is definitely the scariest part of this.
That massive device matrix is exactly why I'm looking for user feedback right now—I really need to see how it holds up on older phones and where it actually breaks.
But even with that headache, I still think it's worth it over just building another wrapper app that calls a server.
The engineering is the impressive part (real-time Magenta on-device, no cloud, is genuinely hard), but the honest read on your numbers: you've built a technical achievement in search of a reason to buy. 2 sales in 3 weeks isn't a pricing problem, it's positioning. "AI music generator on iPhone" competes head-on with Suno and Udio, and you've already conceded you lose on quality. So a buyer's first thought is "why not just use Suno," and the page has no answer.
But you do have one, you've buried it as an apology. "Offline, no server, runs entirely on-device" is your whole product, and you're framing it as a limitation. Flip it. The people who care that it's fully offline aren't Suno's audience at all. They're a different buyer: privacy-focused users, people generating on a plane or off-grid, artists who don't want prompts sent to someone's server, on-device-AI tinkerers. For them "offline" isn't a compromise, it's the entire reason to buy, and Suno structurally can't offer it.
So the wedge isn't "AI music on your phone" (you lose), it's "the only music generator that never touches a server" (you win, uncontested). Same app, different buyer, and one where your battery/speed tradeoffs are acceptable because they came for the offline-ness, not the speed.
That gap between what you built and how it's framed is what I spend my time on, I'm part of the team building Hivemind, an AI strategy copilot. Happy to go deeper on the offline-first repositioning if useful. Either way, stop apologizing for the offline part, it's the only thing that makes you uncopyable.
Thanks! Did you get that same vibe from looking at the website, too? I might need to make some changes there if so