iXnote 26.005

iXnote 26.005 is live on the App Store, and this is the biggest change I've shipped since FollowAlong launched. The short version: FollowAlong now runs the latest open-source AI models entirely on your device, and transcripts are kept for 7 days. Behind that is a real decision about which models to build on. I've been running this build for a couple of weeks and I'm genuinely pleased with it, and now I want to hear how it lands for you. Here's what it means for you, why I made the call, and everything that shipped.

What this means for you

Better meetings, immediately. The new engines hear more accurately and draft sharper cards. Names, terms and decisions that the old stack fumbled now come through, so the cards FollowAlong surfaces mid-conversation are more often the right ones, and the ones it drafts need less editing.

It works when you need it. A meeting assistant gets one chance. This release adds a readiness gate, self-healing audio capture and a live health ticker, so that when you tap record, it records, and when something is wrong, you can see what and fix it in seconds instead of discovering an empty transcript an hour later.

Your hardware keeps up longer. If your device supports Apple Intelligence, FollowAlong runs at full quality. No second tier, no feature you're locked out of because you didn't buy this year's phone.

It gets better on my schedule, not once a year. Owning the model stack means that when a better open model lands in the 4B class, and several will over the next year, I can test it and ship it to you in a regular update. FollowAlong's intelligence is no longer pinned to an annual OS cycle.

Your privacy is structural, not a promise. Everything runs on your device. Transcripts live on your devices and in your private iCloud for 7 days, then they're gone. The cards you chose to keep are the record. I never see any of it, because there is no server to send it to.

Why I chose to run my own models

When FollowAlong first shipped, it ran on Apple Intelligence: Apple's built-in speech transcription and their on-device foundation model. That was a deliberate bet. The models were free, already on the device, and I intended to ride the improvements I assumed would come with each iOS release.

Then I spent time with the early iOS 27 betas, and I started to question that assumption.

I want to be careful here, because these are early betas and the final release doesn't land until September. I'm not going to characterise their performance, and I remain genuinely open to Apple's transcription and model family improving before then. What I can say is that the thing my original bet depended on, steady and predictable improvement I could plan a product around, isn't something I feel I can count on right now.

The real issue is cadence and visibility. Apple appears to update its model catalog once a year, publishes no benchmarks, and shares no roadmap. The open-source world will likely produce several meaningful improvements in the 4B-parameter class over the next year, and running my own models means I can adopt them as they land. An annual, opaque release cycle is a hard foundation to build a business on. A gentle nudge to Apple, from someone who wants them to win: published benchmarks and a faster cadence than once a year would go a long way in AI, where a year is a decade.

I also wanted the widest possible install base. Apple's larger, better on-device model carries a considerably higher hardware bar than the Apple Intelligence baseline. I didn't want to gate FollowAlong to only the people carrying the most advanced hardware, so I drew my line at the Apple Intelligence minimum itself: an M1-class chip and 8 GB of RAM. Qwen, quantized and run through Apple's MLX framework, fits comfortably inside that budget. That gives FollowAlong a significantly larger install pool, and one shared model across iPhone, iPad and Mac instead of different answers on different devices.

None of this is a rejection of Apple's platform work. MLX is excellent, and these model choices are never permanent. I'm always reviewing and refining them, which was the plan from the start. If the final iOS 27 lands transcription and on-device models on par with Parakeet and Qwen, I'd happily shrink the app's footprint and lean on Apple's built-in tools instead. I'm holding that decision until the full public release in September. Until then, FollowAlong works on any device that supports Apple Intelligence, full stop.

Everything in 26.005

New engines, fully on device. FollowAlong's transcription now runs on NVIDIA Parakeet (CC BY 4.0) through Core ML, and its understanding runs on Qwen3-4B (Apache 2.0), quantized to 4-bit through MLX. I didn't swap on a hunch: I built an offline replay rig that feeds recorded meeting transcripts through candidate pipelines side by side, and ran bake-offs for weeks before committing. Nothing changed about the privacy posture: no audio or text ever leaves the device, there is no account and no server.

One-time model download. Two models totalling about 2.7 GB can't live inside an app binary. They arrive as Apple-Hosted Background Assets the first time you open the FollowAlong panel, then the app warms the transcription model immediately so your first recording doesn't pay the compile cost.

A readiness gate on the record button. Recording only becomes available once the models are on disk, warmed, and smoke-tested. You can no longer tap record into a 14-second model compile that misses your opening sentence.

Self-healing audio capture. If the audio stream goes quiet, the capture stack rebuilds itself; if the system resets media services mid-meeting, same thing. Background audio keeps a meeting transcribing when you lock the screen or switch apps.

A live health ticker. One line above the transcript: live words when things are healthy, plain guidance when they're not.

Seven-day transcripts. Transcripts are kept for 7 days and then deleted automatically. Cards you accepted are your own notes and are never touched. Transcript export (Pro) covers the last 7 days.

A recording acknowledgment. Before FollowAlong records for the first time, you confirm with a button press that getting permission from the people you record is your responsibility. One time, synced across your devices.

An honest setup card. Devices that can't run the models, like an older iPhone or an Intel Mac, no longer get offered a download that could never work. The setup card says up front that FollowAlong needs newer hardware, and the rest of iXnote works normally.

Fixes. Bulk meeting import no longer links every imported meeting to today's daily card, connection counts no longer resurrect deleted cards after a replace-import, and a launch deadlock on large graphs is gone.

What's next

The Shortcuts story is designed and building: a small set of composable actions so you can automate cards, connections and daily notes from the Shortcuts app. And FollowAlong keeps getting smarter about your graph, because the thesis behind all of this hasn't changed: the best assistant is one that runs entirely in your pocket and answers to no one but you.

That said, I've been living inside this build for two weeks, and there's only so much one person's testing can tell you. So the part I'm actually looking forward to is you using it. If FollowAlong catches something in your next meeting that the old version would have missed, that's the whole point.

iXnote keeps your cards on your devices and in your private iCloud, end-to-end encrypted. I never see your data, because it never touches my servers.

Update from the App Store, and become a better thinker, not a better note taker.