Ollama Pull Fails with "max retries exceeded" or EOF: How I Get Downloads to Finish

Ollama Pull Fails with "max retries exceeded" or EOF: How I Get Downloads to Finish

ollama pull deepseek-r1:14b crawls to 96%, then: "Error: max retries exceeded: EOF". A 9 GB download dying near the end is the most annoying failure in the local-LLM workflow. Here is my checklist, in the order that solves it fastest.

First: just re-run the pull

Related: Midjourney vs DALL-E vs Flux: A Practical Image-Tool Comparison Framework →

Related: Ollama Error "model requires more system memory": How I Fixed It in 10 Minutes →

Ollama resumes from completed blobs. Re-running is not wasted time, and on a flaky connection two or three attempts genuinely finish the job:

until ollama pull deepseek-r1:14b; do echo "retrying..."; sleep 5; done

That loop is now permanently in my shell history. It rescued the 14B download after four attempts.

Fix 1 — Lower the MTU (VPN and some ISPs)

Related: 9 Free AI Coding Tools Every Developer Should Try in 2026 →

Related: Ollama Answers Get Cut Off Mid-Sentence: num_ctx vs num_predict Explained →

Large packets silently dropped mid-transfer produce exactly this EOF. On my VPN:

ip link show eth0 | grep mtu
sudo ip link set dev eth0 mtu 1400

If that fixes it, make it permanent in your VPN profile instead of per boot.

Fix 2 — Change DNS

Related: Sora 2 Review: OpenAI's Video Model Is Finally Useful for Real Work →

Related: Ollama "connection refused on 127.0.0.1:11434": The 5 Causes I Have Actually Hit →

registry.ollama.ai resolves through a CDN. My ISP resolver kept sending me to a failing edge node.

# /etc/resolv.conf (WSL2: disable generation first in /etc/wsl.conf)
nameserver 1.1.1.1
nameserver 8.8.8.8

Fix 3 — Teach the daemon about your proxy

Related: How to Clone Your Voice with AI in 2026 (Free and Paid Options) →

Related: Ollama Running Slow? 7 Fixes That Actually Worked on My Machine →

Exporting the variables in your shell does nothing — the daemon is a separate process:

sudo systemctl edit ollama
# [Service]
# Environment="HTTPS_PROXY=http://proxy:3128"
sudo systemctl daemon-reload && sudo systemctl restart ollama

Fix 4 — Clear a corrupted partial blob

Related: DeepSeek-R1 Shows Its <think> Tags in the Output — Here Is How I Strip Them →

If one specific model always dies at the same percentage, its partial blob is bad:

ls ~/.ollama/models/blobs | grep partial
rm ~/.ollama/models/blobs/*-partial*
ollama pull deepseek-r1:14b

On macOS the same path applies; on Windows it is C:\Users\<you>\.ollama\models.

Fix 5 — Check free disk before blaming the network

Ollama needs the full model size free, plus temp space. A "download" failure that is really a disk failure looks identical in the log. I hit this at 96% with 6 GB left for an 8.9 GB model — see moving the Ollama models directory.

My working setup

# ~/.bashrc
alias opull='f(){ until ollama pull "$1"; do sleep 5; done; }; f'

Plus MTU 1400 on the VPN interface and Cloudflare DNS. Downloads have not failed since.

FAQ

Does re-pulling waste bandwidth? No. Completed layers are content-addressed and skipped.

Can I copy models between machines? Yes — copy the whole ~/.ollama/models directory. Faster than re-downloading on a slow line.

Where do I choose which model to pull? Sizes and quantization trade-offs are in the local LLM guide; runner comparisons in Ollama vs LM Studio vs Jan and the ultimate AI tools guide.