Ollama 'model requires more system memory (X GiB) than is available (Y GiB)' (HTTP 500) or out-of-memory when loading a model: how do I reduce memory use?

Cause: Ollama checks whether a model's weights plus its context (KV cache) fit in available GPU/system memory before loading. The check fails when the model is too large for the machine, when other loaded models already occupy memory, or when the context length and parallelism make the KV cache large. Ollama's FAQ notes that required memory scales with OLLAMA_NUM_PARALLEL multiplied by the context length, and that models stay loaded for 5 minutes by default. In containers or LXC the reported "available" figure can also be wrong; Ollama's issue tracker has reports of false insufficient-memory errors there (for example when free rather than available memory is read), so compare with what the host really has. Fix (best first): 1. Run ollama ps to see what is loaded, its size and whether it is on GPU or CPU, then free memory with ollama stop <model>. Set OLLAMA_KEEP_ALIVE to a short value (for example 0 or 1m) so idle models unload sooner, and consider lowering OLLAMA_MAX_LOADED_MODELS. 2. Lower the context window: the default is 4096 tokens, and a larger OLLAMA_CONTEXT_LENGTH, num_ctx in a request or a Modelfile raises memory use. Keep OLLAMA_NUM_PARALLEL at 1 unless you need concurrency. 3. Enable OLLAMA_FLASH_ATTENTION=1 (auto-used where supported) and set OLLAMA_KV_CACHE_TYPE to q8_0 (about half the cache memory of f16) or q4_0 (about a quarter); this is global and can affect quality. 4. Pick a smaller model or a more heavily quantized tag (for example a smaller parameter count or lower-bit variant) that fits. 5. In Docker/LXC, check the container memory limit and that the container sees the real memory; raise the limit rather than ignoring the check. Look at the server logs (journalctl -u ollama on systemd Linux) for the exact numbers. Verify: ollama run <model> loads without the error, and ollama ps shows the model with the expected processor split (100% GPU, or a CPU/GPU split you accept). Sources: https://docs.ollama.com/faq https://docs.ollama.com/troubleshooting