Error messages

timed out waiting for llama-server to start

timed out waiting for llama-server to start

You ask for a model, the spinner turns for a long time, and then this comes back (often with a bit more text after it):

Error: timed out waiting for llama-server to start - 

Versions before 0.30.0 said timed out waiting for llama runner to start instead, sometimes with a load percentage in the middle, and that's the wording in the 77 GitHub issues that quote it (as of September 26, 2026). Same check, new name for the helper.

What Ollama was waiting for

Ollama runs each model in a separate helper program called llama-server and polls it until the model is loaded. We read that loop in Ollama 0.34.4. It's a stall timer: every time the helper shows loading activity, the deadline moves forward again. So a huge model can take twenty minutes to load without tripping it, as long as it keeps moving. The error means five whole minutes passed with no progress at all.

If the helper had crashed you'd get a different error, llama-server process has terminated. Here it was still running, just stuck.

Two different endings

Look at what follows the message, because the code has two exits with nearly the same text.

• - and then a line (or nothing) means the five-minute stall timer ran out. The text after the dash is the last error-looking line the helper printed, when there was one.
• : context canceled means the client went away. Whatever asked for the model (a script with its own timeout, a chat front end, a closed terminal) dropped the connection, and Ollama abandons a load nobody's waiting for. The server log says "client connection closed before llama-server finished loading, aborting load". A collaborator explained that in #4131: "if the client cancels the connection before the model finishes loading, we abort the load."

Why loads stall

Issue #4131 is the long thread on this (45 comments, starting with 34B and 70B models on a pair of P40 cards). The collaborator's replies keep coming back to the same few causes. Reading one user's log, he wrote that "we made zero progress loading in 5 minutes and gave up," then asked about slow storage where the models live and whether the system was paging (shuffling memory out to disk because RAM ran out). For someone loading a model of about 150 GB into 128 GB of RAM, his answer was that the system "is likely paging heavily and stalled."

So the usual story is a model too big for memory, or model files on a slow disk or network share, or both.

What to do

Watch the machine during a load with top, free or Task Manager. If memory is full and the disk is busy, the model doesn't fit, and a smaller size tag or a smaller quantization (a more compressed version of the same model) is the real fix.

If the model does fit but loads slowly (a network drive, an old hard disk), give it longer. OLLAMA_LOAD_TIMEOUT is read by the server, not the command line, and takes a duration like 10m or a plain number of seconds. Zero or a negative value turns the stall check off. Set it wherever the server gets its environment (the systemd service on Linux, for instance), then restart Ollama.

The same collaborator also suggested turning off memory-mapped loading for slow loads, with "options": {"use_mmap": false} in an API request. That was advice from mid-2024, so treat it as something to try, not a known fix.

For the context canceled ending, look at the client instead. Raise its timeout, or load the model first with a plain ollama run and let later requests find it already in memory.