This landed on Hacker News four days ago with 513 upvotes and almost no coverage outside of developer circles. It deserves a wider audience.
An indie developer going by Arthur-Ficial published a tool called apfel. It wraps Apple's FoundationModels framework — the on-device language model that ships with every Apple Silicon Mac running Apple Intelligence — and exposes it three ways: a command-line tool, an interactive chat interface, and an OpenAI-compatible HTTP server at localhost:11434.
What that means in plain language: if your team runs on Apple Silicon Macs and has Apple Intelligence enabled, you already have a running AI model. It's sitting there right now. It costs nothing. It runs entirely offline. And until this week, there was no way to actually use it for anything useful outside of Siri.
What Apple hid and why it matters
Apple ships a 3 billion parameter language model baked into every Mac with Apple Silicon. It powers some Apple Intelligence features, but Apple deliberately kept the FoundationModels framework locked behind system-level access. You couldn't pipe a document through it. You couldn't script against it. You couldn't use it in any workflow that Siri wasn't already part of.
apfel changes that by wrapping the framework in interfaces developers and non-developers can actually use.
The model runs entirely on the Neural Engine — the dedicated AI chip built into every M-series processor. No internet connection. No API calls. No data ever leaving the machine. When you feed it a document, that text is processed locally and disappears. There's no server log. Apple collects nothing from on-device inference.
For a 15-person nonprofit that processes donor information, client case files, or grant proposals: this is the first AI tool that a cautious board or IT policy can say yes to without a lengthy review process.
The three things apfel actually gives you
A CLI you can script. echo "Summarize this:" | apfel works. Pipe any file through it. Feed it meeting notes, draft emails, policy docs. Use it in shell scripts and automation workflows. Exit codes work correctly, so it integrates with existing tooling without fuss.
An OpenAI-compatible HTTP server. This is the part that made developers excited. apfel starts a local server on port 11434 — the same port Ollama uses — that speaks the OpenAI API format. Anything that can talk to GPT-4 can talk to this instead, just by changing the base URL to http://localhost:11434. Every internal tool, every agent framework, every workflow your team has already built: it drops in as a replacement with no code changes.
Tool calling support. The model can call functions defined as JSON schemas. That means agent-style workflows — where the AI decides to look something up, format a result, or trigger an action — are supported out of the box. This isn't a toy. It's a real foundation for building lightweight automation.
The honest limitations
The context window is 4,096 tokens. That's roughly 3,000 words. Long documents need to be chunked. If you're processing a 50-page grant report in one shot, this isn't the right tool for that task — a 16GB Mac mini with Gemma 4 E4B (via Ollama) handles long documents much better.
3B parameters is small. You're not getting GPT-4-level reasoning here. For tasks like summarization, draft emails, basic Q&A against short documents, classification, and light writing assistance, it performs well. For complex multi-step reasoning or domain-specific technical questions, bigger models will outperform it.
It requires Apple Silicon and Apple Intelligence to be enabled. Intel Macs are out. Any team with a mixed fleet of hardware — some Apple, some Windows — needs a different solution for the Windows side.
And it's maintained by one developer, published four days ago, with no commercial backing. It's early. The API surface may change.
Who this is for right now
The NGO or public sector team that can't use cloud AI. If your data governance policy says nothing leaves the building — client data, sensitive casework, protected health information, anything subject to PIPEDA or sector-specific regulations — cloud AI tools have been off the table. apfel gives you an on-device alternative that requires no exceptions to your data handling policies. Especially if your team already runs on MacBooks.
The small team that's been sending documents to ChatGPT and hoping for the best. A lot of small orgs are doing this. Nobody has formalized a policy because nobody at the top has asked yet. When someone does ask, you want to have an answer ready. A private, on-device tool is a much easier conversation than explaining why your sensitive files have been hitting OpenAI's servers for the past year.
The team building internal tools. If you have a developer (or someone adjacent to one) building lightweight internal tooling — automating intake forms, processing uploaded files, generating draft responses — the OpenAI-compatible API means they can prototype against the free on-device model and switch to a more capable model later if needed. Same code, different base URL.
How to set it up
Install requires Swift and the Apple Command Line Tools with the current SDK. Detailed install steps are in the apfel docs. Once it's running, the HTTP server launch is one command:
apfel serve
From there, point any OpenAI-compatible client at http://localhost:11434 with model name apfel. It's available to any app on the machine.
For teams that want to use this across multiple users rather than installing on each machine, the setup gets a bit more involved — you'd need to run it on a shared Mac on the network and point clients at that IP instead of localhost. Worth doing, but takes someone who knows their way around the command line.
The bigger picture
Two weeks ago we covered Gemma 4's Apache 2.0 license as a reason to look seriously at self-hosted AI. The trend line is the same here: free, local, private AI is becoming accessible to teams with no dedicated AI budget and no tolerance for cloud risk.
The difference with apfel is that it requires nothing new. No model download. No new hardware. No configuration from scratch. The model is already on your devices. You just didn't have a door.
That changes the ROI calculation for small orgs. You're not asking the board to approve new hardware or a new software subscription. You're asking to turn on something that's already there.
The orgs that move on this in the next few weeks — before the tool matures, before IT policies adapt, before competitors figure out it exists — are the ones who'll have a working private AI workflow while everyone else is still in vendor evaluation.
We've been watching this project since it hit HN and we're already testing it with a couple of clients who have strict data handling requirements. If your team runs on Apple Silicon and you want help figuring out whether this fits your workflow — or how to build something on top of it — let's talk.