Skip to content

Troubleshooting

Symptom → fix. Work top to bottom: the common causes are listed first.

Out of memory / crash when loading a model

Symptom: ollama run fails, the model starts then dies, or the machine staggers.

Fixes:

  1. The model is too big for the memory. Run the size formula and step down — a 14B at ≈ 8.4 GB that won't load becomes an 8B at ≈ 4.8 GB that will. Smaller models also answer faster on the same hardware.
  2. Use heavier quantization. Q4_K_M (≈ 0.6 bytes/param) instead of Q8_0 (≈ 1.0) nearly halves the footprint: an 8B model drops from ≈ 8 GB to ≈ 4.8 GB.
  3. Free memory: ollama ps shows loaded models — /bye out of chats you aren't using so Ollama unloads them. Close GPU-hungry programs (games, video editors).
  4. Leave headroom. The model file, the conversation cache, and your OS share the pool — "exactly fits" doesn't fit.

Slow answers

Symptom: words trickle out; a simple question takes forever.

Fixes:

  1. Check what's doing the work. CPU-only machines are much slower than GPUs at this — if you're on a CPU, small models (1B–3B) are the realistic choice.
  2. Use a smaller model. Less arithmetic per word means faster answers on any hardware.
  3. Close competing programs. Anything else using the CPU or GPU steals speed.
  4. Long conversations slow down too — the conversation memory grows with length. Starting a fresh chat is a legitimate fix.

No absolute speed numbers here on purpose: tokens-per-second depends on your exact setup. Compare relatively — "faster than before" is the test that matters.

Model won't download / pull fails

Symptom: ollama pull errors out or stalls.

Fixes:

  1. You need the internet for the first download — offline comes after the model is on your disk.
  2. Check disk space. A 4.8 GB model needs 4.8 GB free, plus working room.
  3. A corrupted partial download: ollama rm <model>, then ollama pull <model> again for a clean copy.

API: connection refused

Symptom: curl http://localhost:11434/api/tags fails.

Fixes:

  1. Ollama isn't running. Start any ollama run <model> — the API server starts with it — then retry the curl.
  2. Wrong address. On the same machine it's localhost; from another machine it's the host's LAN address (see More Than One Machine), and OLLAMA_HOST must be set on the host.
  3. A firewall on the host machine may be blocking port 11434 — check your OS firewall settings.

Other machine can't reach the host

Symptom: http://localhost:11434 works on the host, but http://<host-address>:11434/api/tags fails elsewhere.

Fixes:

  1. OLLAMA_HOST wasn't set before Ollama started. Set it, fully quit Ollama, start it again.
  2. You're using the wrong address — use the host's LAN address (like 192.168.1.50), not localhost, from the other machine.
  3. Host firewall blocking port 11434 (see above).

Weird, rambling, or wrong answers

Symptom: the model ignores instructions or makes things up.

Fixes:

  1. Simplify the prompt — short, specific, one thing at a time. Small models need simple instructions.
  2. Check you grabbed an instruct model, not a base one (see Talking to Models).
  3. Lower the temperature for factual tasks (see Talking to Models).
  4. Accept the size limit: a 1B model will not reason like an 8B one. If the task needs more, the answer is a bigger model, not a better prompt.

Next: Glossary