Ollama model downloads keep going backwards: progress drops (for example from 60% back to 10%) and the total size shrinks during `ollama pull`

发生场景

Source(s): https://github.com/ollama/ollama/issues/8484 Signal: 32 reactions on the issue, plus 38 on the comment confirming the workaround. Status: reported working by source, not reproduced by us (workaround only; root fix not confirmed; still reported in 2026).

问题

Large model pulls stall, then lose progress and start parts again, on several networks.

报错原文（已脱敏/规范化）

(symptom, no error text) ollama pull progress drops back (for example from 60% to 10%) and the total size shrinks

不是这个问题时 / 边界

The pull fails with an HTTP error such as pull model manifest: 500 (that was a registry outage). With client2, one user's browser-based /api/pull started failing on a CORS preflight.

根因

When a download part stalls for a few seconds, the client drops it and starts that part again. A maintainer also pointed to ISPs that interfere with one registry server.

修复步骤

1. When the speed drops, press Ctrl+C within a few seconds and run the same ollama pull again. The download resumes where it was. Many users confirmed this.
2. Or try the newer client: run the server with OLLAMA_EXPERIMENT=client2 (one user also set OLLAMA_REGISTRY_MAXSTREAMS=1). Two users confirmed their downloads then finished.

如何验证

The pull reaches 100% and ollama list shows the model.