Error messages

Error: Post "http://127.0.0.1:11434/api/chat": EOF

Error: Post "http://127.0.0.1:11434/api/chat": EOF

You run a model and the command gives up before the model writes anything:

Error: Post "http://127.0.0.1:11434/api/chat": EOF

Older versions and other apps hit /api/generate instead, so you'll also see Error: Post "http://127.0.0.1:11434/api/generate": EOF. As of October 1, 2026 the ollama/ollama repo had 30 issues quoting the chat form and 40 quoting the generate form, most of them from 2023 and 2024.

What the message means

The ollama command is a client. It sends your prompt over HTTP to the Ollama server, which listens on port 11434. We read the client in Ollama 0.35.0: if that request fails it hands back Go's own error untouched, and Go formats it as the method, the quoted address and the cause. The cause here is EOF, end of file, which means the connection closed before any HTTP reply arrived.

So the server took the request and went away. A contributor answered exactly this in #2555 (February 2024): the message "just means the server had an issue", and in his own case it had been a segfault, a crash from bad memory access. The client can't say more because it never got an answer.

That's the difference from unexpected EOF, which shows up when a reply had started streaming and then stopped partway. A bare EOF after Post means nothing came back at all.

Finding the actual crash

The reason lives in the server log. On Linux with systemd that's journalctl -u ollama, on a Mac it's ~/.ollama/logs/server.log, and on Windows it's under %LOCALAPPDATA%\Ollama, per Ollama's troubleshooting page. Look at the last lines before the server restarted.

If they're unhelpful, turn on debug output. Running by hand, start it with OLLAMA_DEBUG=1 ollama serve. As a systemd service, add Environment="OLLAMA_DEBUG=1" under [Service], then systemctl daemon-reload and restart it (the steps the contributor gave in #2555).

Here's one we can point at. Collaborator rick-github reproduced the error in #12089 (August 2025) by sending an image to an imported vision model. The CLI printed the bare EOF, and the log ended with ggml-cpu/ops.cpp:6930: fatal error, which he traced to a tensor stored in a type that function couldn't handle.

Causes people found in the threads

• Not enough memory. In #2555 one reporter had about 1.5 GB free on a 2 GB server, and the contributor's verdict was "you won't be able to run much".
• A model file or import the backend can't handle, like the vision merge above.
• CPU or GPU support problems. #2555 also covers a missing AVX2 instruction set and an integrated GPU detection bug that a collaborator fixed for 0.1.29.
• A broken install. The reporter in #2123 (January 2024) saw it on every model right after upgrading.

Current versions report a crashed model runner with their own message, "model runner has unexpectedly stopped", so if you're seeing the bare EOF, check ollama -v as well.

The pull variant: "pulling manifest" then Error: EOF

A shorter Error: EOF right after pulling manifest turns up in 2026 reports. In #16917 it started when the reporter switched Wi-Fi mid-download and kept coming back on every retry. In #17816 one model tag failed this way for 18 hours while another tag of the same model pulled fine. Nobody from the Ollama team gave a cause in either thread beyond asking for server logs, so restart the server, check your network and try a specific tag.