Fix 10 common local AI problems

1. Not enough memory

Symptom: the app reports out of memory, closes during loading, or the whole computer becomes sluggish.

Try: close other large apps, unload other models, choose a smaller complete model and reduce context. In an Ollama chat, /set parameter num_ctx 2048 sets a smaller context for the next request. This is a starting value, not a fit guarantee.

Check: load the model again and send one short message. Use the RAM guide if it still fails.

2. Replies are very slow

Symptom: the model loads, but words arrive slowly.

Try: wait for the first load, shorten the prompt and use a smaller model. ollama ps shows whether the loaded model uses CPU, GPU or both. Native hbay uses the CPU. For GPU setup, follow the runner's current hardware instructions.

Check: start a new chat and ask for a one-sentence reply; avoid comparing different prompt lengths.

3. Wrong or unsupported file type

Symptom: invalid format, unsupported architecture or no model appears in the loader.

Try: select a complete model supported by the app. Do not rename .safetensors to .gguf. Check that you did not choose a tokenizer, adapter, image projection file or only one shard. Update the runner through its official download route if the architecture needs newer support.

Check: load a small model from the runner's own catalog first.

4. The network blocks a download

Symptom: timeout, proxy error, certificate error, or a download stuck before it starts.

Try: finish any Wi-Fi sign-in page, retry on a connection you are allowed to use, or ask your network administrator about the blocked host. On managed networks, use their approved proxy and certificate setup. Ollama documents HTTPS_PROXY; native hbay also supports system certificate trust and custom CA paths.

Check: retry the same download. Keep certificate checks enabled.

5. Not enough disk space

Symptom: no space left, or a download stops near the end.

Try: check free disk space separately from RAM. Make room for the model and any temporary or imported copy. Remove an unwanted model through the app after checking you no longer need it.

Check: retry the download and let its verification finish.

6. A checksum does not match

Symptom: SHA-256 mismatch, corrupt download or verification failure.

Try: confirm you are comparing the hash for the exact file and version. Download that file again from the current source. Do not ignore a mismatch or use a hash from a different quantization.

Check: all 64 SHA-256 characters must match before loading.

7. Command not found

Symptom: the terminal cannot find hbay, ollama or lms.

Try: finish the runner's installation and open a new terminal. For hbay's default Unix install, use export PATH="$HOME/.local/bin:$PATH" in this session. Complete Ollama's CLI setup on macOS. LM Studio's graphical chat does not require lms.

Check: run hbay --version or ollama --version.

8. Cannot connect, or address already in use

Symptom: Ollama cannot reach its server, or a second ollama serve fails.

Try: open the Ollama app. On Linux, if no Ollama service is running, start ollama serve in another terminal. If its port is already in use by Ollama, use the running instance. Do not stop an unknown process just to free a port.

Check: run ollama list.

9. The installer will not open

Symptom: wrong architecture, missing Linux library, signature refusal or an OS version error.

Try: compare your operating system and CPU with the official requirements. Use the published build for that machine. Re-download a damaged installer and check its signature or hash. For native hbay, check OpenSSL 3+ first.

Check: open the supported app and read its version. Use another supported runner if your machine is outside its requirements.

10. The answer is nonsense or ignores you

Symptom: repeated text, strange role labels, or a confident but wrong answer.

Try: start a fresh chat with a short prompt. Choose a chat or instruct model and use its supported chat template. A model imported with the wrong template may not respond properly.

Check: try the same harmless prompt with a small model from the runner's catalog. Compare factual claims with a reliable source; a working model can still be wrong.

Sources and checks

Checked 2026-10-17. Install requirements and availability can change; the linked source has current details.

Read as Markdown