Hours Saved Now Reads On Gemma 4 12B

August 16, 2026

The nightly Hours Saved analysis — the local model that turns each conversation-day into expert hours, a profession, a satisfaction label, and a dollar figure — now runs on Gemma 4 12B. Same privacy rules as before: transcripts stay on the box, tool I/O is stripped, nothing is sent to a cloud scorer. The estimates just get a stronger reader.

What you can do

• Keep using Settings → Account → Hours Saved as before. Totals, the per-day ledger, editable hourly rates, and the 90-day backfill are unchanged.
• Get new days scored by Gemma 4 12B. The 02:00 UTC batch and on-demand Refresh / Scan today now call gemma4:12b on the local Ollama runtime instead of gemma4:e4b.
• Keep the same conservative math. Gemma still reports pure human billable hours; the 50% AI-flex handicap, agent-runtime subtraction, and satisfaction chip are applied in code, not by the model.

Where this shows up

You open Hours Saved after a heavy week and the new conversation cards read more like the work you actually did — the profession call and the one-line summary have more room to be right, because the analyst is no longer the small e4b.

You already trust the page because the numbers never left your machine. That part did not change. Only the local model behind the nightly pass did.

Try it

This one lives in the app, not the chat box:

• Go to Settings → Account → Hours Saved and check yesterday's cards after the next 02:00 UTC run.
• Hit Refresh to score any still-unanalyzed days in the last 90 days on 12B.
• Open Hourly rates if you want the dollar column to match your market — the model upgrade does not reprice anything by itself.

Heads up

• Closed days already scored stay as they are. A row computed after the UTC day closed is final, including receipts written on e4b since launch. Refresh fills missing days; it does not rewrite finished ones.
• Today is still provisional. Mid-day Scan today rows recompute after the day closes, now on 12B.
• Still an estimate. A stronger local model is still a judgment of your transcript, then halved. Treat the output as a defensible lower bound, not an invoice.