Offline, Always: Latent v0.4.1 Beta Is Live

The note isn't an afterthought. For most residents — especially in general surgery — documentation is a significant chunk of the working day, often stretching into the evening after the clinical work is already done. A typical surgical resident might complete eight to fifteen patient encounters in a day. Each one requires some form of written documentation. The mental cost of translating what happened in an exam room into a structured, billable note accumulates in a way that's hard to describe unless you've been doing it for years.

The tools available for this are, in practice, not very good. Voice-to-text works for dictation but doesn't produce structure. Cloud-based ambient scribes exist, but they come with consent requirements, connectivity assumptions, and institutional-approval timelines that make them unavailable on most floors. The gap between what's theoretically possible and what a resident can actually use on a Tuesday morning is wide.

The instinct when building a transcription tool for healthcare is to build it in the cloud. Cloud transcription is fast, inexpensive, and accurate. It's also a PHI problem, a hospital-policy problem, and in some environments a structural impossibility — not because of regulatory caution, but because surgical floors often have poor connectivity, operate inside institutional networks that block external traffic, or fall under policies that prevent health information from leaving the hospital's infrastructure.

Latent was designed offline from the start, not as a compliance checkbox but because that's the only version that works without asking anyone for permission. The transcription pipeline runs entirely on your Mac using locally-hosted models through Ollama. No audio leaves the device. No session is stored remotely. The note is generated locally, ready to be copied into your EMR the moment the encounter ends.

This also means the tool is available to anyone with a compatible Mac, regardless of institutional policy, network access, or vendor approval status. That scope — personal, local, immediately usable — is the point.

The March 5th release addressed two friction points that surfaced repeatedly during early beta feedback. The first was interrupted sessions. The previous pipeline didn't handle mid-session crashes or manual stops cleanly — the local transcription layer would sometimes fail to export from a partial recording, leaving you with nothing. That's fixed. The pipeline now checkpoints continuously and exports cleanly from any interruption point.

The second was first-run setup. Getting Ollama installed and a language model pulled locally is not complicated for developers, but it's genuinely unfamiliar territory for most clinicians. The v0.4.1 setup flow now includes explicit Ollama health checks before attempting first-run model downloads, so you can see exactly what state the environment is in before the tool tries to proceed. Clearer setup feedback sounds minor until you're running through it at 9 PM after a long call day and you just need the thing to work.

Latent requires macOS 12 or later, 16 GB RAM, 10 GB of available disk space for local model storage, and a locally-installed instance of Ollama. Beta access is currently available by email request — not to manufacture scarcity, but because the goal at this stage is to gather feedback from actual clinical workflows before broadening onboarding.

If you're a resident or attending working in a context where documentation load is high and cloud-based tools aren't an option, this is the beta worth testing.

The roadmap beyond v0.4.1 is focused on encounter templates for common note types, improved speaker segmentation for multi-provider encounters, and cleaner export formatting that maps more naturally to the note structures clinicians actually use. The offline model infrastructure is stable, so the remaining work is mostly at the surface: making the transcript-to-note step feel less like a raw dump and more like a note that happens to have written itself.

The longer-term goal is for Latent to handle the full documentation cycle for a clinical shift — not a notes assistant you have to manage, but infrastructure that disappears into the workflow.