Ollama Problems and Fixes: The Errors Users Hit Most, Solved
Ollama fails with a specific error string, and that string decides the fix. Match the most common errors to their fixes, with the exact commands.
By Galactic News Editorial · Published 2026-11-10T16:22:44+00:00 · Updated 2026-11-10T16:22:44+00:00
Ollama Problems and Fixes: The Errors Users Hit Most, Solved
Ollama usually fails with a specific error string, and that string decides the fix. This guide matches the errors users hit most to their fixes, with the exact commands to run. Start with the three diagnostic commands below. They solve most cases in minutes.
Start here: three commands that diagnose most problems
Run these before you search anything. Each one answers a different question.
- ollama --version confirms the binary exists and shows the installed version. If the command is not found, Ollama is not installed or not on your PATH.
- curl http://localhost:11434/api/tags checks whether the server answers. A JSON list of models means the server is up. A connection refused error means it is not running.
- ollama ps shows which models are loaded and where they run. The Processor column says GPU, CPU, or a split between the two. This one command explains most slow responses.
For deeper detail, restart the server with debug logging: OLLAMA_DEBUG=1 ollama serve. The log almost always names the cause. On Linux with the official installer, the server runs as a systemd service, so check it with systemctl status ollama and read recent logs with journalctl -u ollama.
Ollama problems and fixes: GPU not detected
This is the most common Ollama problem. The model loads, but it runs on the CPU at a fraction of GPU speed. The first check is outside Ollama. Run nvidia-smi for NVIDIA cards or rocm-smi for AMD. If the system cannot see the GPU, Ollama cannot use it. Update the GPU driver next. Ollama's documentation sets a minimum driver version for current releases, so an old driver is a real cause, not a guess.
Confirm the diagnosis with ollama ps. If the Processor column says CPU, the model fell back. Check the server log for the reason. On Windows with WSL2, the common trap is installing the Linux CUDA driver inside the distro. WSL2 needs the Windows-side driver instead. The Linux driver package breaks GPU access.
Model will not load: out of memory
An out-of-memory error means RAM or video memory ran out. Two different failures wear the same phrase. The CUDA allocator failing is a video memory problem. The kernel OOM killer is a system RAM problem. The log tells you which one.
Fix it in this order. First, pull a smaller model or a smaller quantisation of the same model. Second, stop models you are not using with ollama stop. Third, check how many models can stay loaded at once with the OLLAMA_MAX_LOADED_MODELS setting. Fourth, check disk space. Model files are large, and a full disk fails a pull in confusing ways. Models live in ~/.ollama/models on macOS, /usr/share/ollama/.ollama/models on Linux with the official installer, and C:\Users\%username%\.ollama\models on Windows. Point OLLAMA_MODELS at another drive when the system disk is tight.
llama runner process has terminated
This is the most reported Ollama failure, and it is not one bug. The runner is the child process that serves the model. When it dies before answering a health check, Ollama prints this message. The exit status printed after the colon is the real diagnosis. Read it before changing anything.
Common causes include running out of memory, a corrupt model file, and a GPU driver problem. Re-pull the model if the file may be corrupt. Free memory and try a smaller model if the exit status points at memory. Run with OLLAMA_DEBUG=1 ollama serve to see the runner's own error line.
Connection refused on port 11434
Ollama serves its API on port 11434 by default. A connection refused error means no server is listening there. The server may not be running, or it may be bound to another address. Check the service first with systemctl status ollama on Linux. Start it if it is stopped. Then check OLLAMA_HOST, the setting that controls the bind address. The default is 127.0.0.1:11434, so only the same machine can connect.
Do not fix this by exposing the port to the internet. The local API has no authentication. Anyone who can reach it can use your GPU and your model library. Keep OLLAMA_HOST on loopback and reach it through an SSH tunnel or a reverse proxy with a password.
Model not found and digest mismatch
A "model not found" error with a 404 usually means the tag is wrong. Check the exact name with ollama list. A common mistake is mixing up the local API and the cloud service. A model name ending in :cloud is a hosted model, not a local one. It leaves your machine and may use paid credits.
A digest mismatch means one cached blob is corrupt. Re-pulling reuses the bad blob and fails the same way. Delete the corrupt blob from the models directory and pull again. The same fix applies when a pull fails halfway and every retry fails identically.
Slow responses and short context
Slow inference almost always means CPU fallback, covered above. If the GPU is in use and answers still crawl, check the context length. Ollama's FAQ states a 4096-token default context. Long prompts get truncated without an error. Raise it with OLLAMA_CONTEXT_LENGTH on the server, or set num_ctx per request, and watch memory climb with it.
Also check the keep-alive window. Models unload after five minutes idle by default. The next call after that reloads the model from disk, which looks like a hang. Raise OLLAMA_KEEP_ALIVE or send a warm-up request on a schedule. A 403 Forbidden error from a browser app is different: it is a cross-origin rejection. Add the origin with OLLAMA_ORIGINS.
Quick reference: error, cause, first fix
| Error you see | Likely cause | First fix |
|---|---|---|
| llama runner process has terminated | Runner died; exit status is the diagnosis | Read the exit code, then check memory and logs |
| out of memory | Video memory or system RAM exhausted | Pull a smaller model; stop unused models |
| Connection refused on 11434 | Server not running or bound elsewhere | systemctl status ollama; check OLLAMA_HOST |
| Model not found (404) | Wrong tag, or a :cloud model treated as local | ollama list; check the exact tag |
| Digest mismatch | Corrupt cached blob | Delete the bad blob, pull again |
| GPU not used, very slow | Driver missing or WSL2 driver mistake | nvidia-smi; fix the driver; check ollama ps |
| 403 Forbidden | Cross-origin rejection | Set OLLAMA_ORIGINS for the app |
| OLLAMA_MODELS ignored | Variable set in shell, not in the service | Set it in the systemd override or launchctl |
Bottom line
Most Ollama problems are one of four things: the server is not running, the GPU is not visible, memory ran out, or the wrong URL is in use. Check the server with curl, check the GPU with ollama ps, and read the log with debug logging on. Fix the cause the error names, not the symptom you feel. And keep port 11434 on loopback unless a password-protected proxy stands in front of it.