Who we are
We are a small, remote group of engineers and designers who have each spent long enough on someone else's cloud to know it's not the only way. Between us there's work at open-source runtimes, a couple of model teams, and a lot of years of building native apps that people actually keep on their phones.
We don't have a CEO-of-the-month photo page. We don't have a 60-slide deck on synergy. What we have is a shared conviction that the best AI you'll use in a given week is the one sitting an inch from your face, ready before you finish the thought — and that conviction is enough to shape every decision we make.
We're not a startup chasing an exit. We're builders who got tired of watching the most transformative technology of our generation get locked behind API keys and monthly invoices. So we did something about it.
The quiet revolution
There's a version of the AI future where a handful of companies own the conversation. You type into their box, your words travel to their servers, their model thinks on their hardware, and the answer comes back — metered, logged, and billed. Your context becomes their training data. Your creativity becomes their moat.
We reject that future.
The open-source community has already proven that powerful models don't need to live behind a paywall. Llama, Qwen, Gemma, Mistral, DeepSeek — these aren't toys. They're capable, general-purpose models that researchers and engineers around the world have poured real work into, and then given away. That generosity deserves software that takes it seriously.
Nimbus8 is that software. We take the best open models, verify them on real hardware, and put them in your pocket. No middleman. No toll booth. No permission slip. The revolution isn't loud — it's a phone that thinks for itself.
Open source, open weights
We don't train proprietary models. We don't believe in them. The open-weights ecosystem is moving faster than any single lab, and the models it produces are good enough to be useful and small enough to run on a phone. That's the sweet spot, and we live in it.
Our runtime wraps MLX, llama.cpp, and Core ML — all open-source projects maintained by communities we respect and contribute back to. The model catalog pulls directly from Hugging Face, the largest open model registry in the world. When you install a model in Nimbus8, you're downloading it from the people who made it.
We believe open source isn't just a licensing strategy — it's a statement about who gets to participate in the future of intelligence. Everyone should. That's the whole point.
Own your models
When you download a model in Nimbus8, it lives on your device. It's a file on your disk, like a photo or a song. You can keep it forever. You can delete it whenever you want. You can install five models or fifty. Nobody can revoke access, change the terms, or sunset the service.
This is what ownership looks like in the age of AI. Not a subscription that gives you permission to use someone else's computer — but a model that runs on your silicon, responds to your prompts, and answers to nobody but you.
No API key. No usage cap. No "we've updated our pricing." The model is yours. The conversation is yours. The compute is yours. That's the deal, and it doesn't change.
Privacy by construction
Most privacy policies are apologies in advance. Ours is three sentences long because there's nothing to apologize for.
Nimbus8 doesn't collect analytics — not even anonymous ones. It doesn't phone home. It doesn't sync your chats to a server. It doesn't run your prompts through a moderation proxy. The network is off by default, and the only time it turns on is when you explicitly ask to download a model or fetch a web search result.
This isn't a feature we bolted on. It's a consequence of the architecture. When inference runs on-device and data never leaves the sandbox, privacy isn't a policy — it's physics. There's nothing to leak because there's nowhere for it to go.
Our mission
To put real, useful AI on the device you already own, without attaching it to a subscription, a login, or a pipe to someone else's server.
The incumbents will tell you that the future of AI is a rented conversation, billed by the token, with your context stored indefinitely on machines you don't control. We disagree. We think the future of AI looks a lot like the future of the camera: capable, quiet, local, and so tightly integrated that you stop thinking about where the compute happens.
Nimbus8 is our attempt to ship that future one iPhone at a time.
Principles
- Local by default. The network is off until you turn it on. Models live on disk, inference runs on your silicon, chats stay in the app's sandbox.
- No accounts. You should not need to prove who you are to talk to software you own.
- Open weights, curated. We don't train proprietary models — we pick, tune, and verify the open ones. The catalog is opinionated; the choice is yours.
- Quiet surfaces. No streaks, no nudges, no "You haven't talked to Claude today." The app should be a place to think, not a game to play.
- Privacy without ceremony. Compliance checklists are a symptom, not the goal. If nothing leaves the device, a lot of the ceremony melts away.
How Nimbus8 got made
The first prototype was a pair of weekend hacks: a Swift wrapper around llama.cpp that streamed tokens into a chat view, and a tiny launcher for picking which model to run. It was bad and slow and we kept using it anyway, because every query felt like ours again.
A year later, the "bad and slow" parts have given way to a proper runtime — MLX for Apple Silicon, GGUF via llama.cpp for broader coverage, Core ML for vision and audio — and the single chat view has grown into eight focused modules named after different kinds of weather. The through-line hasn't changed: we want the software you open when you don't want to think about the internet.
Community
Nimbus8 exists because of the open-source community. Every model in the catalog was trained by researchers who chose to share their work. Every runtime we wrap was built by engineers who chose to publish their code. We're standing on the shoulders of thousands of people who decided that the future of AI should be open, and we take that seriously.
We contribute upstream. When we find a bug in llama.cpp or hit an edge case in MLX, we file it and fix it. When we build tooling that could help others, we open-source it. The ecosystem gave us everything we needed to build Nimbus8, and the least we can do is give back.
Questions about Nimbus8? Email support@driftrail.com or visit support.
What we don't do
- We don't collect analytics. Not even anonymous ones.
- We don't sell your data. (It's not on any server of ours to sell.)
- We don't run a chat log through a moderation proxy. The only code that sees your conversation is the model you picked and the iOS text view.
- We don't ship features we wouldn't use ourselves for a week first.
- We don't charge per token.
Press & brand
A press kit with the Nimbus8 logo, module marks, and screenshots will be available the day the app is live on the App Store. Journalists and reviewers can reach us at the address below in the meantime and we'll be happy to get you early access.
Brand basics, in the meantime: the wordmark is Nimbus8 (one word, capital N, the 8 is a numeral; never spaced or spelled out). The cloud mark is always the smiling face. The primary palette is called Vanilla Wood and that's not negotiable.
Contact
Press, partnership, or "hey, how does this thing actually work" questions: support@driftrail.com. We read it and we answer within a few days. There's no support AI on the other end; just the team.