ollama / ollama

What This Repo Claims

Get up and running with large language models locally with one command. No API key. No cloud. No recurring cost.

171k stars. Released v0.22.1 on May 3, 2026. One of the most popular AI infrastructure projects on GitHub.

The core promise: ollama run llama3.2 — that's it. A local LLM responding to queries in seconds, with a REST API at localhost:11434 that any application can call.

What I Tested

Environment:

• macOS, MacBook Pro 14-inch M4 Pro, 24GB RAM
• Ollama version 0.15.4 (installed 9 months prior)
• Models: llama3.2:latest (2.0GB), deepseek-r1:1.5b (1.1GB)
• No API key used at any point

Note: Testing was conducted on the main development machine, not an isolated test user. Ollama installs as a macOS desktop application and requires system-level installation.

Test 1: Single command query

ollama run llama3.2 "What is a vector database? Answer in one sentence."

Output:

A vector database is a type of database that stores and manages data as dense vectors, 
or multi-dimensional arrays, allowing for efficient storage, similarity search, 
and clustering capabilities for large collections of text documents, 
images, or other data points.

One command. No setup beyond initial install. Correct answer.

Test 2: REST API

curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2",
  "messages": [{"role": "user", "content": "Say exactly: Hello from Ollama"}],
  "stream": false
}'

Response:

{
  "model": "llama3.2",
  "message": {"role": "assistant", "content": "Hello from OLLAMA."},
  "done": true,
  "total_duration": 382298875
}

REST API working immediately. Response in 382ms. JSON structure is OpenAI-compatible.

Test 3: Model switching — same API, different model

curl http://localhost:11434/api/chat -d '{
  "model": "deepseek-r1:1.5b",
  "messages": [{"role": "user", "content": "Say exactly: Hello from DeepSeek"}],
  "stream": false
}'

Response:

"Hello! I'm DeepSeek-R1, an artificial intelligence assistant created by DeepSeek..."

Same endpoint. Same request structure. Only the model field changed. The API switching claim holds.

Findings

Finding 1: Model instruction-following varies

llama3.2 followed the instruction exactly ("Hello from OLLAMA."). deepseek-r1:1.5b ignored the instruction and gave its standard introduction instead.

This is model behavior, not an Ollama failure. Ollama correctly ran both models — what the model does with your prompt is the model's responsibility. Worth knowing when choosing which model to run locally.

Finding 2: Cold start load time

• llama3.2 response time: 382ms (model already in memory)
• deepseek-r1:1.5b response time: 8,710ms total — of which 8,206ms was model loading from disk

Once loaded, deepseek-r1 inference was 242ms — faster than llama3.2. The cold start penalty is significant for the first request but disappears on subsequent calls.

Finding 3: Version tested vs current

Tested on 0.15.4. Current release is 0.22.1 (May 3, 2026). Seven minor versions behind. The core API behavior tested here is stable across versions — the REST endpoint and CLI interface have not changed. New versions add model support and performance improvements.

What I Did Not Test

• Fresh install from scratch (testing on existing install)
• ollama pull to download a new model
• OpenAI-compatible endpoint (/v1/chat/completions)
• Streaming responses
• GPU acceleration (M4 Pro uses Metal — not explicitly tested)
• Windows and Linux installs

Verdict: Solid

The core claim holds completely. One command runs a local LLM. The REST API works immediately with no configuration. Model switching requires changing only the model field.

171k stars. 52 million downloads per month as of Q1 2026. The attention is deserved.

The strategic value for developers: once Ollama is installed, any application can call localhost:11434 with the same interface as a cloud LLM provider — with zero API cost and zero data leaving your machine. This is the unlock for fully local AI demos and prototypes.

Worth installing. One command to start.