Ollama Model Cheatsheet

Q: What's the difference between Ollama and running models directly?

Ollama handles downloading, quantization, GPU acceleration, and API serving automatically. A single `ollama run` replaces 10+ manual steps of PyTorch/HuggingFace setup, CUDA version management, and format conversion.

Q: How do I know which model my computer can run?

Check RAM. 8B models need ~6GB, 4B needs ~3GB, 1B needs ~1GB. Quantized versions (q4/q8) need even less. Your server with 1.8GB RAM + 4GB Swap can handle up to 3B models.

Q: Which model is best for Chinese conversation?

Qwen3 (Alibaba) is the top choice. qwen3:4b outperforms many 7B models at Chinese tasks, and qwen3:8b approaches GPT-3.5 level. DeepSeek also handles Chinese well but focuses more on reasoning.

Q: How to integrate Ollama with external applications?

Ollama provides a REST API on localhost:11434 that's compatible with the OpenAI API format. Use curl directly, or the OpenAI Python/JS SDK with a custom base_url to connect any tool that supports OpenAI's API.

Q: What are embedding models used for?

Embedding models convert text into vectors for semantic search, RAG (Retrieval-Augmented Generation), and text clustering. They don't generate dialogue — pair them with a vector database like Chroma or Qdrant.

Q: Model downloads are slow, what can I do?

Ollama downloads from GitHub Releases by default. Set a mirror via environment variable or use a proxy. For popular models, you can also download GGUF files from HuggingFace and import via Modelfile.

Ollama is one of the most popular tools for running LLMs locally, supporting hundreds of open-source models with one-command deployment. This cheatsheet covers model parameters, memory requirements, use-case recommendations, and complete CLI/API commands.

Updated: 2026-07-17·15 commands

Model Reference Table

🧠 General / Reasoning

Model

Sizes

Context

Highlights

Best For

qwen3	0.6B / 1.7B / 4B / 8B / 14B / 30B / 32B / 235B	128K	Alibaba's latest, thinking mode + tool use, strong multilingual	⭐Chinese tasks
llama3.1	8B / 70B / 405B	128K	Meta flagship, excellent tool calling	General English
llama3.2	1B / 3B	128K	Lightweight & fast	Edge devices
deepseek-r1	1.5B / 7B / 8B / 14B / 32B / 70B / 671B	128K	⭐Best reasoning, CoT, near o3 level	Math/logic/code
gemma3	270M / 1B / 4B / 12B / 27B	128K (4B+)	Google's latest, single GPU capable	Lightweight/vision
kimi-k2.5	-	128K	Multimodal + agentic	Multimodal agents
mistral	7B	32K	Classic v0.3, tool calling	General chat
mixtral	8x7B / 8x22B	32K	MoE, equivalent to 45B/141B	High-value mid-range

👁️ Vision / Multimodal

Model

Sizes

Context

Highlights

qwen3-vl	2B / 4B / 8B / 30B / 32B / 235B	128K	⭐Best vision-language, tools+thinking
gemma3 (4B+)	4B / 12B / 27B	128K	⭐Single-GPU vision, 140+ languages
llava	7B / 13B / 34B	4K	Classic vision-language
llama3.2-vision	11B / 90B	128K	Meta official vision
qwen2.5-vl	3B / 7B / 32B / 72B	32K	Tongyi Qianwen vision, strong Chinese
deepseek-ocr	3B	-	⭐OCR specialist, efficient text extraction
minicpm-v	8B	8K	Efficient vision-language
granite3.2-vision	2B	128K	IBM document understanding
mistral-small3.1	24B	128K	Mistral vision edition

📐 Embedding

Model

Size

Highlights

nomic-embed-text	137M	⭐Most popular, 78.6M pulls
mxbai-embed-large	335M	High accuracy
qwen3-embedding	0.6B / 4B / 8B	Alibaba latest, strong Chinese
all-minilm	22M / 33M	⭐Ultra lightweight
snowflake-arctic-embed	22M~335M	Multi-size options
bge-large	335M	BAAI, good Chinese
embeddinggemma	300M	Google

💻 Code-Specialized

Model

Sizes

Highlights

qwen2.5-coder	1.5B / 7B / 14B / 32B	Alibaba code specialist
codegemma	2B / 7B	Google code
codellama	7B / 13B / 34B / 70B	Meta code
deepseek-coder	1.3B / 6.7B / 33B	DeepSeek coding
starcoder2	3B / 7B / 15B	Multi-language code
laguna-xs.2	33B (MoE, 3B active)	⭐Local agentic coding

---

💾 Memory Requirements

``bash # Model Size → Approx RAM Needed (quantized) 270M~1B → 1GB RAM # Raspberry Pi / phone 3B~4B → 3GB RAM # ⭐Old laptop 7B~8B → 6GB RAM # ⭐Standard home PC 14B → 10GB RAM # High-end laptop 30B~32B → 20GB RAM # Workstation 70B → 40GB RAM # Dual GPU 235B~405B → 140GB RAM # Data center 671B → 400GB RAM # Cluster / H100`

> Your server (1.8GB RAM + 4GB Swap): Max 1B~3B models (qwen3:0.6b / gemma3:1b / llama3.2:1b / deepseek-r1:1.5b)

---

`🏆 Use-Case Picks`

What you need

Best Model

Alternative


Chinese chat qwen3:8b qwen3:4b
English chat llama3.1:8b gemma3:12b
Math/reasoning deepseek-r1:14b qwen3:8b
Coding assistant qwen2.5-coder:14b deepseek-coder:6.7b
Image understanding gemma3:12b qwen3-vl:8b
OCR deepseek-ocr:3b qwen2.5-vl:7b
Embedding/RAG nomic-embed-text bge-large
Ultra lightweight gemma3:1b-qat llama3.2:1b
Tool/function calling qwen3:8b llama3.1:8b
Low-end/server qwen3:0.6b deepseek-r1:1.5b

---
🔧 CLI Commands

`bash # ─── Basic ─── ollama pull deepseek-r1:14b # Download model (specific version) ollama run qwen3:8b # Interactive chat ollama run qwen3:8b "Hello" # Single-turn Q&A ollama list # List downloaded models ollama ps # List running models ollama stop qwen3:8b # Stop a model ollama rm qwen3:8b # Delete a model

# ─── Model Management ─── ollama show deepseek-r1:14b # View model details ollama cp qwen3:8b my-qwen # Clone a model ollama push myuser/my-model # Push to registry

# ─── Custom Models ─── # Create a Modelfile cat > Modelfile << 'EOF' FROM qwen3:8b SYSTEM "You are a concise coding assistant." PARAMETER temperature 0.3 PARAMETER top_p 0.9 EOF

ollama create my-coder -f Modelfile # Build custom model ollama run my-coder # Run custom model

# ─── Environment Variables ─── OLLAMA_HOST=0.0.0.0 # Listen on all interfaces OLLAMA_MODELS=/path/to/models # Model storage path OLLAMA_NUM_PARALLEL=4 # Max concurrent requests OLLAMA_KEEP_ALIVE=5m # Model keep-alive time`

---

`🌐 REST API`

Ollama provides a REST API on localhost:11434 with OpenAI-compatible endpoints.

`bash # Generate text curl http://localhost:11434/api/generate -d '{ "model": "qwen3:8b", "prompt": "What is RAG?", "stream": false }'

# Chat mode (with context) curl http://localhost:11434/api/chat -d '{ "model": "qwen3:8b", "messages": [ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "How to read CSV in Python?"} ], "stream": false }'

# Embeddings curl http://localhost:11434/api/embed -d '{ "model": "nomic-embed-text", "input": "Text to embed" }'

# Streaming (stream: true, default) curl http://localhost:11434/api/generate -d '{ "model": "deepseek-r1:14b", "prompt": "What is 1+1?" }'

# List local models curl http://localhost:11434/api/tags

# Model details curl http://localhost:11434/api/show -d '{"model": "qwen3:8b"}'`

### Python SDK

`python from openai import OpenAI

client = OpenAI( base_url="http://localhost:11434/v1", # Ollama endpoint api_key="ollama" # Not validated by Ollama )

response = client.chat.completions.create( model="qwen3:8b", messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Explain list comprehensions in Python."} ] ) print(response.choices[0].message.content)`

### JavaScript / Node.js

`javascript import OpenAI from 'openai' const client = new OpenAI({ baseURL: 'http://localhost:11434/v1', apiKey: 'ollama' }) const res = await client.chat.completions.create({ model: 'qwen3:8b', messages: [{ role: 'user', content: 'Hello' }], }) console.log(res.choices[0].message.content)``

---

📊 Parameter Counts (Total vs Active)

Model

Total Params

Active Params

Architecture

qwen3:0.6b	0.6B	0.6B	Dense
llama3.2:1b	1B	1B	Dense
deepseek-r1:1.5b	1.5B	1.5B	Dense
qwen3:4b	4B	4B	Dense
gemma3:4b	4B	4B	Dense
qwen3:8b	8B	8B	Dense
llama3.1:8b	8B	8B	Dense
deepseek-r1:8b	8B	8B	Dense (distilled)
gemma3:12b	12B	12B	Dense
qwen3:30b	30B (MoE)	~3B	MoE
qwen3:32b	32B	32B	Dense
mixtral:8x7b	47B (MoE)	~13B	MoE
deepseek-r1:70b	70B	70B	Dense
qwen3:235b	235B (MoE)	~22B	MoE
llama3.1:405b	405B	405B	Dense
deepseek-r1:671b	671B (MoE)	~37B	MoE

---

🚀 Top Models by Pulls

Rank

Model

Pulls

Basic Operations(5)

Command	Level	Example
`ollama pull <model>` Download a model from the registry to local	Basic	`ollama pull llama3.2`
`ollama run <model>` Run a model and enter interactive mode	Basic	`ollama run llama3.2`
`ollama list` List locally available models	Basic	`ollama list`
`ollama rm <model>` Remove a model from local	Basic	`ollama rm llama3.2`
`ollama cp <src> <dst>` Copy a model (create an alias)	Intermediate	`ollama cp llama3.2 my-llama`

Configuration(4)

Command	Level	Example
`ollama show <model>` Show model details (parameters, template, system prompt)	Intermediate	`ollama show llama3.2 --modelfile`
`ollama create <name> -f ./Modelfile` Create a custom model from a Modelfile	Intermediate	`ollama create mymodel -f ./Modelfile`
`ollama push <model>` Push a model to a remote registry	Intermediate	`ollama push mymodel`
`OLLAMA_MODELS=/path/to/models ollama serve` Specify custom model storage directory	Expert	`OLLAMA_MODELS=/data/models ollama serve`

Advanced Operations(6)

Command	Level	Example
`ollama serve` Start the Ollama server (default localhost:11434)	Intermediate	`ollama serve`
`ollama ps` List models currently loaded in memory	Basic	`ollama ps`
`curl http://localhost:11434/api/generate -d '{"model":"llama3.2","prompt":"Hello"}'` Call model via REST API for text generation	Intermediate	`curl http://localhost:11434/api/generate -d '{"model":"llama3.2","prompt":"Hello","stream":false}'`
`curl http://localhost:11434/api/chat -d '{"model":"llama3.2","messages":[{"role":"user","content":"hi"}]}'` Chat via REST API with message history	Intermediate	`curl http://localhost:11434/api/chat -d '{"model":"llama3.2","messages":[{"role":"user","content":"hi"}],"stream":false}'`
`ollama stop <model>` Stop a running model	Basic	`ollama stop llama3.2`
`OLLAMA_HOST=0.0.0.0 ollama serve` Listen on all network interfaces (allow remote access) Ensure firewall protection when enabling remote access	Expert	`OLLAMA_HOST=0.0.0.0 ollama serve`

FAQ

What's the difference between Ollama and running models directly?

How do I know which model my computer can run?

Which model is best for Chinese conversation?

How to integrate Ollama with external applications?

What are embedding models used for?

Model downloads are slow, what can I do?

This cheatsheet is compiled from official tool documentation. Last updated: 2026-07-17.

1	llama3.1	117.2M	93
2	deepseek-r1	89.9M	35
3	nomic-embed-text	78.6M	3
4	llama3.2	76.7M	63
5	gemma3	38.6M	26
6	qwen2.5	34.8M	133
7	qwen3	32.3M	58
8	mistral	31.1M	84
9	gemma2	27.9M	94
10	llama3	24.7M	68
11	llava (vision)	14.3M	98
12	mxbai-embed-large	12.6M	4
13	minicpm-v	5.3M	17
14	llama3.2-vision	4.9M	9
15	qwen3-vl	4.5M	57

Chinese chat	qwen3:8b	qwen3:4b
English chat	llama3.1:8b	gemma3:12b
Math/reasoning	deepseek-r1:14b	qwen3:8b
Coding assistant	qwen2.5-coder:14b	deepseek-coder:6.7b
Image understanding	gemma3:12b	qwen3-vl:8b
OCR	deepseek-ocr:3b	qwen2.5-vl:7b
Embedding/RAG	nomic-embed-text	bge-large
Ultra lightweight	gemma3:1b-qat	llama3.2:1b
Tool/function calling	qwen3:8b	llama3.1:8b
Low-end/server	qwen3:0.6b	deepseek-r1:1.5b