AiOS Dispatch 6

Welcome to the sixth dispatch of AiOS Dispatch, your go-to resource for the latest in AI, language models, IDE updates, and iOS development.

I started writing this on the phone while on my way to Hiroshima, and ended up after finalizing it after coming back from Kyoto. 🌸

A lot has piled up since the last dispatch (usual in the AI world), so here is a quick one!

WWDC 2025

As mentioned in the previous dispatch, I am heading to US this June for WWDC and was lucky enough to secure the ticket too!

I will be around from 7th to 14th of June in Cupertino, so if you are attending the iOS developer's yearly pilgrimage, let me know!

Llama 4 Herd

Meta dropped the Llama 4 family, because who does not love a Saturday surprise? We have got two models right now: Llama 4 Scout and Llama 4 Maverick, with a third, Llama 4 Behemoth, still in the training.

I have been experimenting with these using Together AI and Groq, and they are decent models, but not SOTA according to my initial impressions.

The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation

We’re introducing Llama 4 Scout and Llama 4 Maverick, the first open-weight natively multimodal models with unprecedented context support and our first built using a mixture-of-experts (MoE) architecture.

Llama 4 Scout

First up, Llama 4 Scout. This one is lean with 17 billion active parameters, 16 experts, and a jaw-dropping 10 million token context window. Enough to summarize an entire code base. I doubt the quality of the output anything above 200K tokens, though.

Meta claims it outshines Google's Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 on benchmarks. I threw the codebase (266K tokens) of MLX Swift examples via Together AI, and it was able to output a decent workflow to add local LLM support to your app.

Llama 4 Maverick

Then there is Llama 4 Maverick, the multi-modal one. It is 17 billion active parameters but with 128 experts and 400 billion total parameters. This one got a 1 million token context window which is still massive. It is tuned for chat and creative tasks, and Meta’s brags that it beats GPT-4o and Gemini 2.0 Flash on coding, reasoning, and image tasks.

It is cheaper than proprietary models, with inference costs around $0.19–$0.49 per million tokens. I still think the updated DeepSeek V3 is a better model then this one.

Llama 4 Behemoth

And the tease with Llama 4 Behemoth. Still in training, this one boasts 288 billion active parameters, 16 experts, and nearly 2 trillion total parameters. Meta’s calling it a “teacher” model, distilling its wisdom into Scout and Maverick. No release date but they mention that it is better than GPT-4.5 and Claude 3.7 Sonnet on STEM benchmarks.

I will wait for Behemoth for iOS development tasks but for now, Gemini 2.5 Pro remains my default model for Alex Sidebar and Cursor.

The World of MCP

MCP this, MCP that.

If you have been anywhere near AI discussions in the past few months, you must have heard the acronym thrown around like anything.

But when you see OpenAI and Microsoft jump on the bandwagon, you know it is time to start paying attention.

So, what exactly is MCP? It stands for Model Context Protocol, an open-source standard kicked off by Anthropic in late 2024. A classic example is of being an adapter to connect large language models to external tools, APIs, databases, and more, without custom integrations for every single use case. The protocol standardizes how LLMs talk to these external systems, using a client-server setup where “MCP servers” expose tools and data that the clients can tap into.

Xcode MCP

One MCP that is beneficial for iOS developers is the XcodeBuild MCP. It provides Xcode-related tools for integration with Claude Desktop, and other AI editors.

There is a growing MCP ecosystem but I still feel the friction between the learning curve of setting up MCP servers. The one who can nail the UX for it will benefit loads by bringing it to the majority.

GitHub’s MCP Server

GitHub on boarded the train as well with their official MCP that plugs in into the their APIs.

GitHub - github/github-mcp-server: GitHub’s official MCP Server

GitHub’s official MCP Server. Contribute to github/github-mcp-server development by creating an account on GitHub.

GitHubgithub

Automating PRs, fetching file diffs, or listing issues without leaving your AI tool of choice is another way to reduce friction for iOS development workflow!

Moving Forward

There is a lot to cover in the coming weeks but I feel these were the gist for you this week.

Also, I am heading to the biggest iOS conference after WWDC next week: try! Swift Tokyo

try! Swift Tokyo

Developers from all over the world will gather for tips and tricks and the latest examples of development using Swift. The event will be held for three days from April 9 - 11, 2025, with the aim of sharing our Swift knowledge and skills and collaborating with each other!

try! Swift Tokyo

There is one talk on AI-assisted Swift that I am looking forward to. If you are attending the conference, or stay in Tokyo, let me know and we can meet!