On October 15, Meta released the weights for Muse Spark Open, an 816-billion-parameter model, and showed it running on a desk. Two Mac Studios, each with an M5 Ultra and 512 GB of unified memory, hold the whole thing between them. In Meta's words, "two Mac Studios behave like one supercomputer with a terabyte of unified memory."

Read the announcement as a price list and a pattern shows up. The weights are a free download for many developers. The seven engines Meta names are open-source projects. The Apple tuning ships in MLX, Apple's own framework, which is also free. Almost every layer it takes to run a model this size at home now costs nothing beyond the hardware.

In April we wrote that the M5 Ultra would serve "models that today require a workstation with two H100s." Meta just ran an 816-billion-parameter model on two of them. We built ToolPiper on that bet, and Meta's release raises the question we built it to answer. If the model and the engine are free, where does the value go?

Can you run Muse Spark Open on a Mac?

Yes, with two Mac Studios. Meta says Muse Spark Open runs on two Apple M5 Ultra Mac Studios with 512 GB of unified memory each, using Apple's MLX framework, so the pair holds the entire 816-billion-parameter model.

The model is a mixture of experts. It has 816 billion parameters in total, but only 42 billion of them are active for each token. That split is why a Mac can do this at all. Memory has to hold every parameter, while bandwidth only has to move the active slice. Meta's announcement says the M5 Ultra's 1.2 TB/s of memory bandwidth "is fast enough to read the small slice each token needs."

Size is the first thing people ask about, so here's the arithmetic. At 4 bits per weight, 816 billion parameters come to about 408 GB before any context. At 8 bits it's about 816 GB. The context window reaches 1 million tokens, and long context needs memory of its own. Meta's tested setup is two machines, which is the straight answer to "what Mac do I need?" Our Muse Spark Open page on Model Fit runs the same math against every Mac.

So Muse Spark Open is out of reach for almost every Mac in use today, and that's fine. A model this size marks the ceiling. Meta's own smaller open model, Muse Glimmer 30B, ships an official 4-bit build that Meta says is meant for 24 GB and 32 GB consumer hardware. Everyday work runs on models like that, and they keep improving. Stanford researchers found the best local model answered 71.3% of real-world chat and reasoning queries correctly in 2025, up from 23.2% in 2023 (Intelligence per Watt).

What did Meta give away, and what did it keep?

Meta released the Muse Spark Open weights under its Muse Model License, which has limited commercial terms, and named seven free engines that run them. Paid hosted access comes from seven partners, and Meta's own agent, Muse Code, ran its on-device demos.

Look at where money still changes hands. Hosting is a meter, billed by each partner for every request, and it's the same meter we've argued against since we started. The other layer that still takes real work to build isn't on the price list at all. It's the agent.

Meta didn't demo a chat box. It showed Muse Code shipping real changes to Ice Cubes, an open-source Mastodon client for iOS, and driving DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.

That second demo is the one to study. A capable model can judge which shot is strongest. What it can't do alone is open the project, find the clip on the timeline, move the playhead, apply the crop and check the result. Every one of those steps happens in an app on a particular machine, with that machine's files, windows and permissions.

Why does free software move the value?

Nobody can charge rent on a layer anyone can download, so the value moves to the layers that still take work to build. With open weights and open engines, that leaves hosting for people who rent compute and the software that acts on the machine for people who own it.

Software has done this before. Free web servers didn't end the web business. They pushed it up into the applications built on top. Free compilers didn't end software either. What changed each time was which layer a company could build on.

For local AI, the free layers are now the model and the engine. llama.cpp, the open-source engine behind many local AI apps, appears on Meta's list. So does MLX. Neither one costs anything, and neither one opens a project or clicks a button. A local AI app in 2026 is mostly the part it builds on top, so the useful question about any of them is what it does once the model is loaded.

That's the business we're in. Every open model anyone releases makes the layer above it more capable the day it ships, because tools don't care which model calls them. Meta's release is a gift to that layer, and ToolPiper is that layer.

What does ToolPiper do on your Mac?

ToolPiper runs AI models on your Mac and gives them over 450 tools to act on it, across voice, files, the browser, other apps and system settings. Any AI client that speaks MCP, including Claude Code and Claude Desktop, can call the same tools.

One native app holds the models, the speech, the vision and the tools, so a spoken request can become a finished action on the Mac in front of you. Text models run on llama.cpp, the same engine on Meta's list. MLX powers image, video and speech.

The tools act through the same interfaces you use. ToolPiper finds a button in another app by its role and name in the accessibility tree, clicks it, and reports whether the click took. It drives Chrome through the browser's own debugging protocol. It reads and writes files, arranges windows, changes display and sound settings, and creates calendar events and reminders. You decide which tools run on their own and which ones ask you first.

The models and the tools are separate on purpose. Use a local model for work that should never leave the Mac. Point Claude Code at ToolPiper when you want a frontier model doing the thinking. The tools behave the same either way, because the Mac is the same either way.

Can ToolPiper use Muse Spark Open?

Yes. Meta named seven hosting partners for Muse Spark Open, and six of them, OpenRouter, Together AI, Fireworks AI, Groq, Nebius and Baseten, are built into ToolPiper as providers. Choose Muse Spark Open with your own API key and ToolPiper's tools act on your Mac.

Four of those six have been ToolPiper providers since spring. Nothing had to be built for Meta's release, because a new model arriving at a provider ToolPiper already speaks to is the case it was designed for.

Running it entirely on your own hardware is a different setup. Meta's local path splits the model across two Mac Studios with MLX, and if you own that pair, it's the way to run it. For models that fit on one Mac, ToolPiper runs them locally and your prompts stay on the machine.

Local by default, hosted when you choose

ToolPiper doesn't make you pick a side. Everyday work runs on a local model and stays on your Mac. When a job calls for something bigger, such as Muse Spark Open or a closed frontier model, point the same tools at it and keep working. The tools, the permissions and the Mac stay the same.

The trade is plain. Hosted access sends your prompts to whichever partner serves the model, under that partner's terms, and cloud models through ToolPiper are part of Pro. Local models cost nothing per request, and the Stanford numbers show how much of the daily load they already carry.

What does Meta's release mean for Mac users?

The model and the engine cost nothing for anyone with the hardware, so the software that does the work on your Mac is the choice that matters.

Meta paid to train an 816-billion-parameter model, tuned it for Apple Silicon and gave the weights away. Every Mac owner benefits, whether or not they download it, because it raises what people expect from open models. Our bet is that the next model at this level fits on fewer machines, and the one after that fits on one.

When that happens, the model becomes the part you download once and stop thinking about. What you'll notice is whether your Mac can act on what the model decides. That's the part we build.

Download ToolPiper, load a model, and ask it to do something on your Mac.

This is part of a series on local-first AI on macOS. For the argument behind it, read Why PiperKit Exists.