Meta's Muse Spark Open Puts 816B Open Weights on Two Mac Studios

What Happened: Meta Open-Sources Muse Spark

On October 15, Meta released the weights for Muse Spark Open, calling it "the most capable US open-weight model to date." The numbers behind the claim are genuinely unusual: 816 billion total parameters in a mixture-of-experts architecture with only 42 billion active per token, a 1-million-token context window, and native multimodality across text and images. The weights are live now on Hugging Face, with hosted inference available through Together AI, Fireworks, Baseten, Modal, and OpenRouter at roughly $1.25 per million input tokens and $4.25 per million output tokens from third-party hosts.

But the headline that got developers talking wasn't the parameter count — it was the hardware. Muse Spark Open is the first frontier-class model explicitly optimized to run locally, on Apple silicon, specifically on M5 Ultra Mac Studios. Mark Zuckerberg credited direct engineering work with Apple's hardware chief John Ternus. After years of open-weight models that technically shipped with downloadable weights but practically needed a data center, Meta just moved the goalposts to a desktop.

The Specs: 816B Parameters, 42B Active, 1M Context

Muse Spark Open's architecture is why any of this is feasible. As a sparse mixture-of-experts model, only 42B of its 816B parameters fire on any given token — around 5% of the network. That keeps per-token compute in the range of a mid-size model while retaining the knowledge capacity of a giant one. The million-token context means entire codebases, document archives, or hours of transcripts fit in a single prompt without retrieval gymnastics.

The release also ships with practical tooling rather than just a checkpoint: custom auto-tunable Helion GPU kernels in SGLang delivering 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs, an MLX adapter for Apple silicon that pulls weights straight from Hugging Face, and TorchTitan support for further fine-tuning. It's licensed under the Muse Model License with limited commercial terms — free for many developers, though not a permissive open-source license like Apache 2.0.

The Mac Studio Trick: Two Machines, One Model

How do you run 816B parameters on a desk? Apple's unified memory architecture. The M5 Ultra's integrated GPU can address up to 512 GB of unified RAM per machine — so two Mac Studios linked together can hold the entire model in memory, a terabyte of unified RAM acting like a single supercomputer. The chip's 1.2 TB/s memory bandwidth is fast enough to stream the small active slice each token needs, which is exactly the workload sparse MoE models were designed for.

The catch is cost. Engadget's math: a workable setup — say four 256 GB M5 Ultra Mac Studios — runs roughly $38,000, and the 512 GB configuration isn't even shipping until late October. That's far beyond a hobbyist rig, but for a team currently spending five figures a month on frontier API tokens, a fixed-cost local cluster that never rate-limits, never logs your data, and never changes terms starts to look like infrastructure rather than indulgence. It's the same economics that already pushed developers toward smaller self-hosted models via stacks like Together AI or Fireworks — now available at frontier scale.

Benchmarks: Meta's Open-Weight Answer to China

The geopolitical framing is explicit. Meta has said the US should not let Chinese labs lead on open-source AI — and for most of 2026, they have: DeepSeek, Alibaba's Qwen family, and Zhipu's GLM line (see DeepSeek V3) dominated the open-weight leaderboards. Meta's own benchmarks claim Muse Spark Open outperforms GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning — the two categories that matter most for real tool use in 2026. Independent verification is still arriving, and Meta's evaluation methodology report is worth reading before betting a budget on it. On safety, Meta says the model clears the thresholds of its Superintelligence Scaling Framework, with additional safety post-training applied before release.

Agentic On-Device: Real iOS Commits and Video Editing

The most interesting proof points aren't benchmarks at all — they're agentic workflows Meta ran entirely on-device with Muse Code, its coding agent powered by the new model:

• Shipping real iOS changes to Ice Cubes, the open-source Mastodon client — bug fixes, new features, and a full successful build, with no cloud model in the loop.
• Driving DaVinci Resolve end to end, having the model review footage frame by frame to find the strongest moments and crop them vertically for short-form output.

That's the quiet story here: the frontier of AI tooling in 2026 is agents that operate your actual software — the direction Claude Code and computer-use agents popularized in the cloud. Muse Spark Open demonstrates the same class of capability running on hardware you own. For anyone who has wanted an always-on coding or automation agent without per-task token anxiety or corporate data leaving the building, this is the first credible local path.

What It Means for the AI Tools You Use

Three consequences worth planning for:

• "Open weights" now means runnable, not theoretical. When a US frontier-class model fits on two desktops, self-hosting stops being a hobbyist niche and becomes a serious procurement option for any team with sensitive data or high token volume.
• The local-tool ecosystem will explode. Every coding agent, document assistant, and automation framework that currently requires an API key will race to add local-model backends. Expect a wave of MLX-native tools; inference platforms like Modal and Baseten will host the same weights for teams that don't want the hardware.
• Hybrid stacks become the default architecture. Route hard reasoning to a frontier API, run high-volume and privacy-sensitive work locally. With quality gaps closing, the routing decision becomes economic rather than forced.

Meta's bet is that the future of "personal superintelligence" — their phrase — runs on machines you own. With Muse Spark Open, that idea just got its first frontier-grade proof. The $38,000 question is how fast the hardware curve brings it to the rest of us.