Configuration

All daemon and CLI configuration is read through envconfig/config.go (environment variables) and ~/.ollama/server.json (a small persistent config). The full source of truth is envconfig.AsMap().

Daemon

Variable	Default	Effect
OLLAMA_HOST	127.0.0.1:11434	Bind address (and where the CLI looks for the daemon).
OLLAMA_ORIGINS	localhost / loopback / app://* / file://* / tauri://* / vscode-webview://* / vscode-file://*	CORS allow-list.
OLLAMA_MODELS	~/.ollama/models	Where models live on disk.
OLLAMA_KEEP_ALIVE	5m	How long the scheduler keeps a runner loaded after the last request. Negative means infinite; zero means evict immediately.
OLLAMA_LOAD_TIMEOUT	5m	How long the scheduler waits for a stalled load before giving up. Zero or negative means infinite.
OLLAMA_MAX_LOADED_MODELS	0 (auto: 3 × #GPUs, capped)	Maximum runners loaded at once.
OLLAMA_MAX_QUEUE	512	Pending request queue depth before the scheduler returns 503.
OLLAMA_NUM_PARALLEL	1	Concurrent requests per loaded runner.
OLLAMA_GPU_OVERHEAD	0	Reserve VRAM headroom (bytes) per GPU.
OLLAMA_FLASH_ATTENTION	false	Enable flash attention.
OLLAMA_KV_CACHE_TYPE	(model-specific)	KV cache quantization (e.g., f16, q8_0).
OLLAMA_NEW_ENGINE	false	Prefer the ollamarunner Go-native engine.
OLLAMA_LLM_LIBRARY	autodetect	Force a specific GPU library (cuda_v12, rocm, cpu).
OLLAMA_SCHED_SPREAD	false	Always schedule a model across all GPUs.
OLLAMA_MULTIUSER_CACHE	false	Optimize prompt caching for multi-user scenarios.
OLLAMA_NO_CLOUD	false	Disable cloud features (remote inference + web search).
OLLAMA_REMOTES	ollama.com	Allowed cloud hosts.
OLLAMA_AUTH	false	Enable client/server auth.
OLLAMA_VULKAN	false (Linux/Windows only)	Experimental Vulkan support.
OLLAMA_CONTEXT_LENGTH	0 (auto)	Default context length.
OLLAMA_DEBUG	0	Log level (1/true = DEBUG, 2 = TRACE).
OLLAMA_DEBUG_LOG_REQUESTS	false	Log inference request bodies + replay curl commands to a temp directory.
OLLAMA_NOPRUNE	false	Skip orphan-blob garbage collection on startup.
OLLAMA_NOHISTORY	false	Disable readline history in ollama run.
OLLAMA_EDITOR	$EDITOR fallback	Editor invoked by Ctrl+G in interactive mode.

GPU library selection

Variable	Effect
CUDA_VISIBLE_DEVICES	NVIDIA visible devices.
HIP_VISIBLE_DEVICES	AMD by numeric ID.
ROCR_VISIBLE_DEVICES	AMD by UUID or numeric ID.
GGML_VK_VISIBLE_DEVICES	Vulkan visible devices.
GPU_DEVICE_ORDINAL	AMD by numeric ID.
HSA_OVERRIDE_GFX_VERSION	Override the gfx for AMD GPUs.

These are passed straight through to the runner subprocess by filteredEnv in llm/server.go.

Proxies

HTTP_PROXY, HTTPS_PROXY, NO_PROXY (plus their lowercase variants on non-Windows) are honored by every outbound HTTP call (registry, cloud).

Server config file

~/.ollama/server.json is a small JSON document loaded by envconfig.loadServerConfig (envconfig/config.go). Current fields:

Field	Type	Effect
disable_ollama_cloud	bool	Same as OLLAMA_NO_CLOUD=1.

envconfig.NoCloudSource() returns "none" / "env" / "config" / "both" so callers can explain why cloud is disabled.

Ollama Cloud experiments

OLLAMA_EXPERIMENT (read directly in server/routes.go) is a comma-separated list of experiment names. Currently recognized: client2. Used by handlers that gate behavior behind feature flags.

Runner-side knobs

GGML_* variables are forwarded to the runner subprocess where llama.cpp respects them (e.g., GGML_VK_VISIBLE_DEVICES, GGML_METAL_*).

Where to add a new variable

1. Add a typed accessor in envconfig/config.go (Bool, Uint, Uint64, or String).
2. Add it to envconfig.AsMap() so it shows up in diagnostics and the FAQ scrape.
3. Document it in docs/faq.mdx.
4. Read it via the accessor — never os.Getenv directly elsewhere.