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.
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.
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.