The Hidden Ledger of Ollama and the Multi-Threaded Answer of aria2: The SQLite Override of the Model Directory, the Three-Line Field Test of a 17GB Model, and the Falsification of ollama-pull Acceleration

Two small cases of download and storage closed the same day, both starring downloading and storage — and their shared lesson is that the real behavior of an application lives in its database, not its documentation.

I. The Hidden Ledger of Ollama

The first step of moving the Ollama model directory to drive F was the obvious one: change the user environment variable OLLAMA_MODELS. It did not take — the server log still showed E:\OllamaModels. Digging revealed a hidden behavior: desktop build 0.40.15 stores the model directory setting in its own SQLite database (db.sqlite, the settings.models field) and overrides the environment variable at startup. The fix changed both layers: the variable as a floor, then closing Ollama, backing up the database, updating the settings.models field, and restarting — the server log confirming OLLAMA_MODELS=F:\OllamaModels with all seven models recognized. The database backup was left at backup_20261107; the contents of both drives were compared identical (22 blobs) before the old E-drive directory was declared safely deletable.

II. The Three-Line Field Test of a 17GB Model

The second case was the route selection for a download, field-tested rather than assumed. The master worried 17GB would crawl without a proxy, and three lines of data answered: HuggingFace direct at ~1.8MB/s single-connection with parallel connections useless (per-stream limiting on the official CDN); the hf-mirror.com mirror at ~1.6MB/s single but scaling linearly with parallel connections (8-16 threads reaching 10MB/s+); and ollama pull riding the same official CDN — no acceleration whatsoever, poor resume, no mirror option, falsified. The conclusion: the hf-mirror mirror plus aria2 sixteen connections (about thirty minutes for the whole model), then import into Ollama.

III. Two Victories of Field-Testing

The shared method of the two cases is one line: speak with data, not with legend. The model directory should be decided by the environment variable — field-testing found SQLite overriding it; ollama pull should be the proper road — field-testing found it riding the same slow line as the browser. The 128GB of free space on F was checked, and the 22-blob identity of both drive directories was compared, all measured rather than assumed. On storage and downloading, legends are plentiful — and every one of them can be dismantled by a fifteen-minute field test.