Ollama Cheat Sheet (commands, API, env vars)

A one-page reference for running models locally with Ollama. Tested on Ollama 0.40.13, Windows, CPU-only (no discrete GPU). Chinese version of the site: irubik.cn.

Basic commands

Command	What it does
ollama list	List installed models and their size
ollama pull <model>	Download / update a model
ollama run <model>	Start an interactive chat
ollama ps	Show models currently loaded in memory
ollama stop <model>	Unload a running model
ollama serve	Start the local API server (usually runs automatically)
ollama rm <model>	Delete a model

Useful flags for ollama run: --verbose prints timing statistics after each answer.

Local API

The server listens on http://127.0.0.1:11434 by default.

# generate
curl http://127.0.0.1:11434/api/generate -d "{\"model\":\"qwen2.5:3b\",\"prompt\":\"hello\",\"stream\":false}"

# embeddings
curl http://127.0.0.1:11434/api/embed -d "{\"model\":\"bge-m3\",\"input\":\"some text\"}"

# loaded models and memory usage
curl http://127.0.0.1:11434/api/ps

Timing fields in a /api/generate response:

Field	Meaning
load_duration	Time spent loading the model (nanoseconds)
prompt_eval_count / prompt_eval_duration	Prompt tokens and time
eval_count / eval_duration	Generated tokens and time

Tokens per second = eval_count / (eval_duration / 1e9). Add "keep_alive": 0 to unload the model after the request.

Environment variables

Variable	Purpose
OLLAMA_HOST	Address the server binds to (e.g. 0.0.0.0)
OLLAMA_MODELS	Where model files are stored
OLLAMA_KEEP_ALIVE	How long a model stays loaded after a request
OLLAMA_NUM_PARALLEL	Number of parallel requests
OLLAMA_MAX_LOADED_MODELS	How many models may stay loaded
HTTPS_PROXY	Proxy used when downloading models

Set them before starting ollama serve; on Windows use setx or the system environment settings.

Measured on a CPU-only machine

Intel i5-12400, 32 GB RAM, no discrete GPU, Ollama 0.40.13, one prompt, 80 generated tokens:

Model	Size	Cold start	Speed	Peak memory
qwen2.5 3B	1.9 GB	2.31 s	17.0 tok/s	~2.5 GB
qwen2.5 7B	4.7 GB	5.07 s	7.76 tok/s	~5.5 GB
deepseek-coder 6.7B	3.8 GB	8.33 s	9.16 tok/s	~11.7 GB

All three ran purely on CPU (size_vram was 0). Full write-up (in Chinese): running local models without a GPU.

Embedding models

Model	Output dimension	150-character input
bge-m3	1024	~340 ms
nomic-embed-text	768	~143 ms

Both are fast enough to embed a personal knowledge base on CPU; the bottleneck in a RAG setup is usually generation, not embedding.

Notes

• First answer after loading is always slower; the model stays in memory for a while, so back-to-back questions are faster.
• Memory usage is not the same as model file size — a 3.8 GB model can take 11.7 GB of RAM once loaded.
• If memory is tight, lower the context length or use a smaller model instead of forcing a bigger one.