On October 15, Meta did something the open-source AI world has been waiting years for: it released the weights of a genuinely frontier-class model. **Muse Spark Open** packs 816 billion total parameters (42 billion active per token) into a natively multimodal Mixture-of-Experts model with a **1-million-token context window** — and Meta proved it runs on a desktop. Not a rack of H100s. Two Mac Studios.

How an 816B model fits on a desk

The trick is Apple’s unified memory architecture plus aggressive quantization. Each M5 Ultra Mac Studio exposes up to 512GB of unified memory to its GPU, so two machines behave like one supercomputer with a terabyte of memory. In 4-bit NVFP4 precision the model needs roughly 493GB for weights — it cannot fit on one machine, but it fits on two, connected over Thunderbolt with Apple’s MLX framework doing the orchestration. Apple’s own measurements: about 15–25 tokens per second on two machines, over 36 tokens/sec on four. Slow by data-center standards, usable by human standards.

• ▸816B total / 42B active parameters (256 experts, 8 active per token), 62 layers
• ▸1M-token context; natively multimodal (text, image, document, audio, video in — text out)
• ▸Weights in BF16 and NVFP4 on Hugging Face, SageMaker, Google Cloud, Microsoft Foundry
• ▸API hosts: Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together AI
• ▸License: Muse Model License — free to download, limited commercial use

Meta did not just benchmark it — it shipped real work with it

This is the part that separates the announcement from the usual benchmark theater. Meta paired the model with Muse Code, its AI coding terminal, and ran it end to end on a Mac Studio setup: shipping real iOS changes to Ice Cubes, an open-source Mastodon client — bug fixes, new features, full builds — and driving DaVinci Resolve, with the model reviewing video footage frame by frame to find the strongest moments and crop them vertically. An 816B model doing your video edits on your desk is a sentence that would have sounded absurd two years ago.

The honest caveats

Meta calls it “the most capable US open-weight model to date” — note the careful framing, because that claim rests on Meta’s own evaluations and pointedly excludes China. On Terminal-Bench 2.1, an independent agentic coding benchmark, Muse Spark Open scored 76.4 — behind Kimi K3 (85.0), GLM 5.3 (83.9), and Qwen3.8 (82.0). It leads on some agentic tool-use tests (MCP-Atlas 89.2, top of Meta’s comparison table), but the coding crown among open models still sits in Shenzhen and Beijing. Then there is the hardware asterisk: the 512GB Mac Studio configuration it was tested on **is not on sale yet** — Apple says later this month — and the 256GB model already costs over $10,000. “Runs on your desktop” currently means a desktop budget in the five figures.

Why this matters beyond the spec sheet

Three shifts are hiding inside this launch. First, the **open-vs-closed gap is now a deployment gap, not a capability gap**: a downloadable model can now do long-horizon agentic work that used to require renting a frontier API. Second, **Apple just became AI infrastructure** — Meta’s CEO publicly thanked Apple’s CEO, and MLX is now a serious inference stack. Third, the timing: this landed the same week as Meta’s Muse Wonder and Google’s Gemini 4 Argon going GA. The frontier is fragmenting into ecosystems, and open weights are Meta’s wedge — a model you can own, fine-tune with TorchTitan, and run behind your own firewall is a different product than an API key, even when the weights come with a license leash.

Ozero’s take

I care about this as someone who lives on free tiers and rented APIs: every capability that moves from “rent per token” to “download once” changes who gets to build. An 816B agent-coding model you can self-host kills an entire category of API bills — and for crypto-adjacent builders who need agents touching wallets and signing flows, running the model locally instead of shipping prompts to a third party is not just cheaper, it is a security posture. The benchmark gap with Chinese models is real, but directionally this is the most important open release of the year. Watch what gets built on it in the next 90 days — that will tell you more than any benchmark table.