Free disk space lets you store a model. It does not guarantee enough working memory to run it. Model weights, conversation context and other running programs all compete for resources. A long conversation can fail even when a short prompt worked.
Start with the least disruptive fixes
- Save your work, then close memory-heavy games or applications you do not need. Do not end unknown system processes.
- In Ollama, run
ollama psto see loaded models. Unload an unused one withollama stop MODEL_NAME, substituting its exact name. This unloads it from memory rather than deleting the downloaded file. - Try a smaller model or a more compressed quantization. In LM Studio, the different Q-labelled downloads are quantized variants; smaller downloads can reduce the memory burden but can also reduce answer quality. A model's download size is not its full runtime requirement.
- Reduce the context setting and retry a short prompt. In an Ollama interactive chat,
/set parameter num_ctx 4096is a useful modest test where supported by the model. Lower context means less conversation/document text fits, so it is a trade-off rather than a free performance upgrade.
Check what actually loaded
ollama ps reports whether inference is using CPU, GPU or a mixture. CPU use is not itself a failure, but moving work out of graphics memory can change responsiveness. Confirm your GPU is supported before assuming more system RAM will solve a driver problem.
Stop guessing when failures persist
Record the exact model tag, quantization, context setting, RAM/VRAM and error. Compare a small prompt against the failing workload. Consult the official GPU support and troubleshooting pages before changing drivers or runtime overrides. Do not disable security controls or buy hardware solely from a model's parameter count.
Documentation checked November 14, 2026; no hardware benchmark is claimed.