Frost Window: 30 years of weather and a laptop-sized Gemma tell you what to do outside this week
Description and use
Frost Window uses 30 years of daily low temperature data from Open‑Meteo and a local Gemma 3 4B model to calculate frost risk and generate practical outdoor activity recommendations.
Problem
Gardeners in cold regions need accurate frost risk information to plan planting and outdoor activities, but most apps provide only an average first frost date, hiding the true risk.
Outcome
The tool provides accurate frost risk estimates and actionable outdoor plans, runs entirely offline after initial data fetch, and can be swapped to a smaller model if needed.
Workflow
- 1. User runs
python frost_window.py "Place" --model gemma3:4b - 2. The script queries Open‑Meteo geocoding and historical archive APIs for 30 years of daily lows
- 3. It computes last spring frost and first fall frost for each year and derives typical, early, and 9‑of‑10‑year safe dates
- 4. It formats these statistics into concise fact sentences
- 5. Gemma 3 4B, running locally via Ollama, transforms the facts into a four‑line outdoor plan
Evidence
Frost Window gives you the risk, then gets out of the way:
It pulls 30 years of daily low temperatures for that spot from the free Open‑Meteo archive.
A small open‑weight model, Gemma 3 4B running locally through Ollama, turns those sentences into a four‑line "go do this outside" plan.
The screen part takes about thirty seconds. Everything it suggests happens outdoors.
The 1B model invented places. With gemma3:1b, Seoul got a "Moon Valley" walk that does not exist. Moving to gemma3:4b fixed it.
The code now says "Frost is rare here" when frost occurs in fewer than half the years, instead of pretending a pattern exists.
- Record date: 2026-10-08T03:41:08.343Z