The Definitive Ollama Masterclass

ollama pull gemma4          # download a model
ollama run gemma4           # download if needed, then chat
ollama list                 # list installed models + sizes
ollama ps                   # show what is loaded in GPU/CPU right now
ollama show gemma4          # architecture, parameters, license, context
ollama rm gemma4            # delete a model to free disk
ollama cp gemma4 my-gemma   # copy a model under a new name
ollama serve                # run the API server in the foreground
ollama launch claude        # wire Ollama into Claude Code (also: codex, opencode, droid, copilot-cli)

# Inside the chat REPL:
#   /set parameter num_ctx 8192   -> change context length
#   /show info                    -> model details
#   /bye                          -> exit

# Fit more context on a 12 GB GPU and keep the model hot
export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE=q8_0
export OLLAMA_CONTEXT_LENGTH=8192
export OLLAMA_KEEP_ALIVE=-1
export OLLAMA_MODELS="$HOME/ollama-models"

# On Windows PowerShell, use: setx OLLAMA_FLASH_ATTENTION 1  (etc.), then restart Ollama
ollama serve

# 1) Browse ollama.com/library and note a model name + size tag
# 2) Pull a specific size/quant with a tag after the colon
ollama pull qwen2.5-coder:14b
# 3) Run it
ollama run qwen2.5-coder:14b
# 4) Inspect what you got (context, quant, license)
ollama show qwen2.5-coder:14b

ollama pull gemma4
ollama pull qwen3.5
ollama pull gpt-oss:20b
ollama pull mistral-small3.2
ollama pull granite4.1:8b

ollama pull qwen2.5-coder:14b
ollama pull devstral
ollama pull codestral
ollama pull deepseek-coder-v2

ollama pull deepseek-r1:14b
ollama pull qwq
ollama pull phi4-reasoning

ollama pull qwen3-vl:8b
ollama pull llama3.2-vision
ollama pull moondream

ollama pull nomic-embed-text
ollama pull embeddinggemma
ollama pull mxbai-embed-large

ollama pull gemma3:1b
ollama pull llama3.2:3b
ollama pull phi4-mini
ollama pull smollm2:360m

ollama pull hermes3:8b
ollama pull functiongemma
ollama pull llama3-groq-tool-use

ollama pull mathstral
ollama pull sqlcoder
ollama pull llama-guard3

# Save as "Modelfile" (no extension)
FROM qwen2.5-coder:14b

SYSTEM """You are a senior Shopify + React engineer.
Write production-ready code with error handling.
Vanilla JS only, no jQuery. Scope all CSS under a unique wrapper class.
When unsure, say so and propose a test rather than guessing."""

PARAMETER temperature 0.3
PARAMETER num_ctx 8192

# Build and run it:
#   ollama create shopify-dev -f Modelfile
#   ollama run shopify-dev

# Point FROM at a local GGUF file (or a safetensors directory)
echo 'FROM ./my-model.gguf' > Modelfile
ollama create my-model -f Modelfile
ollama run my-model

# Share it to your namespace on ollama.com
ollama push yourname/my-model

curl http://localhost:11434/api/chat -d '{
  "model": "gemma4",
  "messages": [{ "role": "user", "content": "Why is the sky blue?" }],
  "stream": false
}'

# pip install ollama
from ollama import chat

resp = chat(model='qwen2.5-coder:14b', messages=[
    {'role': 'system', 'content': 'You output only valid JSON.'},
    {'role': 'user', 'content': 'Give me a product with name and price.'}
], format='json')          # force structured JSON output
print(resp.message.content)

# Embeddings for RAG:
from ollama import embed
vec = embed(model='nomic-embed-text', input='chunk of text')['embeddings']

// npm i ollama
import ollama from 'ollama';

const res = await ollama.chat({
  model: 'gemma4',
  messages: [{ role: 'user', content: 'Plan a week of social posts.' }],
});
console.log(res.message.content);

# pip install openai
from openai import OpenAI

client = OpenAI(base_url='http://localhost:11434/v1', api_key='ollama')  # key ignored
resp = client.chat.completions.create(
    model='qwen2.5-coder:14b',
    messages=[{'role': 'user', 'content': 'Explain how Ollama works.'}],
)
print(resp.choices[0].message.content)

# Wire Ollama into a coding agent in one command
ollama launch claude        # Claude Code
ollama launch codex         # Codex
ollama launch opencode      # OpenCode
ollama launch droid         # Droid
ollama launch copilot-cli   # GitHub Copilot CLI

# Turn Ollama into a personal assistant across chat apps
ollama launch openclaw

# ~/.continue/config.yaml — point Continue at your local Ollama
models:
  - name: Qwen Coder (local)
    provider: ollama
    model: qwen2.5-coder:14b
  - name: Autocomplete (local)
    provider: ollama
    model: qwen2.5-coder:1.5b
    roles: [autocomplete]

export OLLAMA_API_BASE=http://127.0.0.1:11434
aider --model ollama/qwen2.5-coder:14b

ollama pull qwen2.5-coder:14b     # chat/edit model
ollama pull qwen2.5-coder:1.5b    # fast autocomplete
# Install the Continue extension, select Ollama, pick the models above.
# Result: tab completion + inline chat, fully local, zero cloud calls.

ollama pull nomic-embed-text     # embedder
ollama pull gemma4               # answerer
# Fastest GUI path: install AnythingLLM or Open WebUI, point at Ollama,
# drop your files in, and ask questions grounded in them.
# Script path (Python):
from ollama import embed, chat
docs = ["...your chunks..."]
vectors = [embed(model='nomic-embed-text', input=d)['embeddings'][0] for d in docs]
# store vectors in any vector DB, retrieve top-k for a query, then:
answer = chat(model='gemma4', messages=[{'role':'user','content':'CONTEXT + QUESTION'}])

ollama pull devstral              # coding-agent model (24B, needs ~24 GB)
ollama launch opencode           # or use Cline in VS Code with Ollama
# Give it a bounded task, review the plan, then let it edit across files.
# On 12 GB, use qwen2.5-coder:14b instead and keep tasks smaller.

ollama pull qwen3-vl:8b          # general vision-language
ollama pull deepseek-ocr:3b      # token-efficient OCR
# CLI: pass an image path in the REPL, or via the API:
from ollama import chat
r = chat(model='qwen3-vl:8b', messages=[{
  'role':'user','content':'Extract the totals from this receipt as JSON.',
  'images':['receipt.jpg']}], format='json')
print(r.message.content)

from ollama import chat
items = open('inbox.txt').read().split('\n---\n')
for i, text in enumerate(items):
    r = chat(model='gemma4', messages=[
      {'role':'system','content':'Classify as [urgent], [fyi], or [archive]. One label only.'},
      {'role':'user','content': text}])
    print(i, r.message.content.strip())

from ollama import chat
r = chat(model='granite4.1:8b', messages=[
  {'role':'system','content':'Return only JSON matching {name, email, company}.'},
  {'role':'user','content':'Jane Doe, jane@acme.co, works at Acme.'}
], format='json')
import json; print(json.loads(r.message.content))

from ollama import chat
def ask(prompt, sensitive=True):
    model = 'qwen2.5-coder:14b' if sensitive else 'qwen3-coder'  # local vs cloud
    return chat(model=model, messages=[{'role':'user','content':prompt}]).message.content
# Sensitive work never leaves the machine; non-sensitive scale routes to Cloud.

# macOS: quit the app, then remove it and the models
rm -rf ~/.ollama
# Linux (script install):
sudo systemctl disable --now ollama
sudo rm /usr/bin/ollama /etc/systemd/system/ollama.service
sudo rm -rf /usr/share/ollama
# Windows: uninstall from Settings > Apps, then delete %USERPROFILE%\.ollama