Error messages

Error: max retries exceeded

Error: max retries exceeded

This one shows up partway through ollama pull (or ollama run for a model you haven't got yet), usually after the progress bar has been crawling for a while:

pulling manifest
Error: max retries exceeded: Get "https://dd20bb891979d25aebc8bec07b2b3bbc.r2.cloudflarestorage.com/ollama/docker/registry/v2/blobs/sha256/...": dial tcp: lookup dd20bb891979d25aebc8bec07b2b3bbc.r2.cloudflarestorage.com: no such host

We read the download code in Ollama 0.34.4. It splits each model file into parts and fetches up to 16 at once. Each part gets six attempts, waiting 1, 2, 4, 8, 16 and then 32 seconds between them. When one part fails all six, Ollama stops the pull and prints max retries exceeded followed by whatever went wrong on that last try. So the first half of the line is always the same, and the useful half is after the colon. It's common: 95 issues in the ollama/ollama repo contain the phrase as of September 26, 2026.

Reading the part after the colon

The long URL in the middle is where the file actually lives. The registry redirects downloads to Cloudflare storage, which is why you see r2.cloudflarestorage.com instead of ollama.com. Skip past it to the end of the line:

• no such host means your machine couldn't look up that storage address. A collaborator on the repo called it "some sort of DNS issue" in #8605 (January 2025), and people in that thread got past it by switching DNS servers.
• EOF, unexpected EOF or connection reset by peer mean the connection was cut mid-transfer. #12624 reports this on large models (small ones got through) behind corporate proxies that inspect encrypted traffic.
• TLS handshake timeout means the secure connection to the storage server never finished setting up, which points at a slow or filtered link.
• A write error naming a -partial file is local: Ollama couldn't write to its models folder. One report (#10072) had the folder on a cloud-storage mount that didn't allow it.

A full disk doesn't end up here. Ollama stops on the first "no space" error without retrying, so you'd see that error on its own.

Slow downloads that stall

A part that moves no data for 30 seconds is restarted, and those restarts don't count against the six attempts. The server log says "stalled; retrying" when that happens. So a slow connection on its own won't trigger this error. It takes six real failures, like dropped connections or failed lookups.

What to do

Run the same pull again first. Finished pieces are kept on disk as -partial files next to the blob, and the next pull reads them and carries on. On a flaky link it can take a few goes.

If it fails straight away every time, test the network path without Ollama. In #8605 a collaborator suggested fetching one model file with curl, following redirects, from https://registry.ollama.ai/v2/library/mistral/blobs/sha256:.... The same collaborator said that if curl also fails when Ollama does, the problem is at the storage end (or on the way to it), and Ollama itself is fine. For no such host, try another DNS server. Behind a company proxy, ask whether it's inspecting or cutting long downloads.

You may find older advice to set OLLAMA_EXPERIMENT=client2. That experimental client has gone: in 0.34.4 the server only logs that it's "no longer available" and asks you to remove the variable.

The same error on ollama push

Uploads use the same six-try limit and the same wording. Push reports in the tracker end with things like http status 502 Bad Gateway or 413 Request Entity Too Large, and those are status codes sent back by the registry itself.