Ollama

“Only Apple can do this” Variously attributed to Tim Cook

Apple introduced Apple Intelligence at WWDC 2024. After waiting almost a year for Apple to, in Craig Federighi’s words, “get it right”, its promise of “AI for the rest of us” feels just as distant as ever.

While we wait for Apple Intelligence to arrive on our devices, something remarkable is already running on our Macs. Think of it as a locavore approach to artificial intelligence: homegrown, sustainable, and available year-round.

This week on NSHipster, we’ll look at how you can use Ollama to run LLMs locally on your Mac — both as an end-user and as a developer.

---

What is Ollama?

Ollama is the easiest way to run large language models on your Mac. You can think of it as “Docker for LLMs” - a way to pull, run, and manage AI models as easily as containers.

Download Ollama with Homebrew or directly from their website. Then pull and run llama3.2 (2GB).

$ brew install --cask ollama
$ ollama run llama3.2
>>> Tell me a joke about Swift programming.
What's a Apple developer's favorite drink? 
The Kool-Aid.

Under the hood, Ollama is powered by llama.cpp. But where llama.cpp provides the engine, Ollama gives you a vehicle you’d actually want to drive — handling all the complexity of model management, optimization, and inference.

Similar to how Dockerfiles define container images, Ollama uses Modelfiles to configure model behavior:

FROM mistral:latest
PARAMETER temperature 0.7
TEMPLATE """
{{- if .First }}
You are a helpful assistant.
{{- end }}

User: {{.Prompt}}
Assistant: """

Ollama uses the Open Container Initiative (OCI) standard to distribute models. Each model is split into layers and described by a manifest, the same approach used by Docker containers:

{
  "mediaType": "application/vnd.oci.image.manifest.v1+json",
  "config": {
    "mediaType": "application/vnd.ollama.image.config.v1+json",
    "digest": "sha256:..."
  },
  "layers": [
    {
      "mediaType": "application/vnd.ollama.image.layer.v1+json",
      "digest": "sha256:...",
      "size": 4019248935
    }
  ]
}

Overall, Ollama’s approach is thoughtful and well-engineered. And best of all, it just works.

What’s the big deal about running models locally?

Jevons paradox states that, as something becomes more efficient, we tend to use more of it, not less.

Having AI on your own device changes everything. When computation becomes essentially free, you start to see intelligence differently.

While frontier models like GPT-4 and Claude are undeniably miraculous, there’s something to be said for the small miracle of running open models locally.

• Privacy: Your data never leaves your device. Essential for working with sensitive information.
• Cost: Run 24/7 without usage meters ticking. No more rationing prompts like ’90s cell phone minutes. Just a fixed, up-front cost for unlimited inference.
• Latency: No network round-trips means faster responses. Your /M\d Mac((Book( Pro| Air)?)|Mini|Studio)/ can easily generate dozens of tokens per second. (Try to keep up!)
• Control: No black-box RLHF or censorship. The AI works for you, not the other way around.
• Reliability: No outages or API quota limits. 100% uptime for your exocortex. Like having Wikipedia on a thumb drive.

Building macOS Apps with Ollama

Ollama also exposes an HTTP API on port 11434 (leetspeak for llama 🦙). This makes it easy to integrate with any programming language or tool.

To that end, we’ve created the Ollama Swift package to help developers integrate Ollama into their apps.

Text Completions

The simplest way to use a language model is to generate text from a prompt:

import Ollama

let client = Client.default
let response = try await client.generate(
    model: "llama3.2",
    prompt: "Tell me a joke about Swift programming.",
    options: ["temperature": 0.7]
)
print(response.response)
// How many Apple engineers does it take to document an API? 
// None - that's what WWDC videos are for.

Chat Completions

For more structured interactions, you can use the chat API to maintain a conversation with multiple messages and different roles:

let initialResponse = try await client.chat(
    model: "llama3.2",
    messages: [
        .system("You are a helpful assistant."),
        .user("What city is Apple located in?")
    ]
)
print(initialResponse.message.content)
// Apple's headquarters, known as the Apple Park campus, is located in Cupertino, California.
// The company was originally founded in Los Altos, California, and later moved to Cupertino in 1997.

let followUp = try await client.chat(
    model: "llama3.2",
    messages: [
        .system("You are a helpful assistant."),
        .user("What city is Apple located in?"),
        .assistant(initialResponse.message.content),
        .user("Please summarize in a single word")
    ]
)
print(followUp.message.content)
// Cupertino

Generating text embeddings

Embeddings convert text into high-dimensional vectors that capture semantic meaning. These vectors can be used to find similar content or perform semantic search.

For example, if you wanted to find documents similar to a user’s query:

let documents: [String] = …

// Convert text into vectors we can compare for similarity
let embeddings = try await client.embeddings(
    model: "nomic-embed-text", 
    texts: documents
)

/// Finds relevant documents
func findRelevantDocuments(
    for query: String, 
    threshold: Float = 0.7, // cutoff for matching, tunable
    limit: Int = 5
) async throws -> [String] {
    // Get embedding for the query
    let [queryEmbedding] = try await client.embeddings(
        model: "llama3.2",
        texts: [query]
    )
 
    // See: https://en.wikipedia.org/wiki/Cosine_similarity
    func cosineSimilarity(_ a: [Float], _ b: [Float]) -> Float {
        let dotProduct = zip(a, b).map(*).reduce(0, +)
        let magnitude = { sqrt($0.map { $0 * $0 }.reduce(0, +)) }
        return dotProduct / (magnitude(a) * magnitude(b))
    }
    
    // Find documents above similarity threshold
    let rankedDocuments = zip(embeddings, documents)
        .map { embedding, document in
            (similarity: cosineSimilarity(embedding, queryEmbedding),
             document: document)
        }
        .filter { $0.similarity >= threshold }
        .sorted { $0.similarity > $1.similarity }
        .prefix(limit)
    
    return rankedDocuments.map(\.document)
}

Building a RAG System

Embeddings really shine when combined with text generation in a RAG (Retrieval Augmented Generation) workflow. Instead of asking the model to generate information from its training data, we can ground its responses in our own documents by:

1. Converting documents into embeddings
2. Finding relevant documents based on the query
3. Using those documents as context for generation

Here’s a simple example:

let query = "What were AAPL's earnings in Q3 2024?"
let relevantDocs = try await findRelevantDocuments(query: query)
let context = """
    Use the following documents to answer the question. 
    If the answer isn't contained in the documents, say so.
    
    Documents:
    \(relevantDocs.joined(separator: "\n---\n"))
    
    Question: \(query)
    """

let response = try await client.generate(
    model: "llama3.2",
    prompt: context
)

---

To summarize: Different models have different capabilities.

• Models like llama3.2 and deepseek-r1 generate text.
  • Some text models have “base” or “instruct” variants, suitable for fine-tuning or chat completion, respectively.
  • Some text models are tuned to support tool use, which let them perform more complex tasks and interact with the outside world.

• Models like llama3.2-vision can take images along with text as inputs.

• Models like nomic-embed-text create numerical vectors that capture semantic meaning.

With Ollama, you get unlimited access to a wealth of these and many more open-source language models.

---

So, what can you build with all of this?
Here’s just one example:

Nominate.app

Nominate is a macOS app that uses Ollama to intelligently rename PDF files based on their contents.

Like many of us striving for a paperless lifestyle, you might find yourself scanning documents only to end up with cryptically-named PDFs like Scan2025-02-03_123456.pdf. Nominate solves this by combining AI with traditional NLP techniques to automatically generate descriptive filenames based on document contents.

The app leverages several technologies we’ve discussed:

• Ollama’s API for content analysis via the ollama-swift package
• Apple’s PDFKit for OCR
• The Natural Language framework for text processing
• Foundation’s DateFormatter for parsing dates

Looking Ahead

“The future is already here – it’s just not evenly distributed yet.” William Gibson

Think about the timelines:

• Apple Intelligence was announced last year.
• Swift came out 10 years ago.
• SwiftUI 6 years ago.

If you wait for Apple to deliver on its promises, you’re going to miss out on the most important technological shift in a generation.

The future is here today. You don’t have to wait. With Ollama, you can start building the next generation of AI-powered apps right now.