The uncomfortable truth about AI chatbots
When you type a question into a mainstream cloud AI assistant, your words travel to a data center. They get processed by a model running on specialized hardware that you don't own, in a building you've never visited, under policies you probably haven't read.
For casual questions, like "what's a good recipe for lentil soup", this trade-off is easy to accept. But that's not how most people actually use AI assistants.
People use them to draft sensitive emails. To brainstorm business ideas they haven't told anyone about yet. To process thoughts they wouldn't say out loud. To write journal entries, therapy reflections, legal strategies, medical questions.
Every one of those conversations goes to someone else's server.
What "private AI" usually means
Most AI companies address privacy concerns with policies: "We don't use your data for training." "Your conversations are encrypted in transit." "You can delete your history."
These are good policies. They're also promises, and promises can change. Terms of service get updated. Companies get acquired. Engineers make mistakes. Data that exists on a server can be subpoenaed, breached, or accidentally exposed.
The safest data is data that never leaves your device in the first place.
A different architecture
That's why we built Nashi.
Nashi runs a large language model directly on your iPhone using llama.cpp. The entire inference pipeline, your prompt going in and the response coming out, happens on the chip in your pocket. No network request is made. No data leaves the device.
This isn't a thin client connecting to a local server. It's not an on-device cache with cloud fallback. The model weights are stored on your phone. The computation happens on your phone. If you put your phone in airplane mode, Nashi works identically.
The trade-offs
Let's be honest about what you give up:
Model size. The models that fit on a phone are smaller than the large models behind cloud chatbots. They're less capable for complex reasoning, less knowledgeable about niche topics, and occasionally less coherent for long outputs. On-device AI is getting better every generation, but right now, cloud models are more capable.
Speed. On-device inference is slower than cloud inference for large models. Your phone is impressive, but it's not a rack of data-centre GPUs. You'll notice the difference for longer outputs.
No internet knowledge. Nashi can't browse the web or access real-time information. It knows what its training data contained, and nothing more.
These are real trade-offs. We think they're worth it for the conversations that matter most: the ones where privacy isn't a nice-to-have, it's the whole point.
Who this is for
Nashi isn't trying to replace cloud assistants for everyone. It's for specific people in specific moments:
Therapists and counselors who use AI to draft session notes or brainstorm treatment approaches, and can't risk client data touching a third-party server.
Lawyers who need to analyze case details or draft arguments without creating a data trail outside their control.
Journalists who use AI to organize source material and can't risk exposing confidential sources.
Anyone who has thoughts they'd share with an AI assistant but not with a corporation. Half-formed ideas, personal reflections, creative experiments, embarrassing questions: the kind of thinking that deserves a space with no audience.
The model keeps getting better
On-device AI models improve every few months. Phones get faster with each chip generation. The models that run on a phone today are meaningfully better than what was possible two years ago.
This means the gap between cloud AI and on-device AI is shrinking. Not disappearing, but shrinking. And for many use cases, on-device models are already good enough.
The question isn't whether on-device AI will be capable enough. It's whether you want to wait for it to catch up while continuing to send your private thoughts to someone else's server.
Try it
Nashi is available on the App Store for a one-time purchase. No account, no subscription, no telemetry. Download a model, start chatting. Your conversations stay on your device, now and always.
If the model isn't capable enough for your needs today, that's fine. Delete the app and your data goes with it. That's the point.