Meta's Muse Spark Open Is an 816B-Parameter Open-Weight Model — and It Runs on Two Mac Studios
Meta released Muse Spark Open on October 15: 816 billion parameters of open-weight AI for coding and agents, tuned to run on a pair of Mac Studios. Here's what matters and what doesn't.
Meta announced Muse Spark Open on October 15, 2026, and the company's framing is not subtle: it calls the model "the most capable U.S. open-weight model to date." The headline specs are 816 billion total parameters with 42 billion active per token, a 1 million-token context window, and tuning aimed at coding and agentic work — extended tasks, multi-step reasoning, software engineering. The model is free to download with a license that allows limited commercial use, and the weights are already on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, with API access through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.
The part Meta really wants you to notice is where it runs. Working with Apple, Meta tuned Muse Spark Open for the Mac Studio, and claims that two 512GB M5 Ultra Mac Studios running the latest MLX framework can hold the entire model. Apple's unified memory architecture lets each machine's integrated GPU address up to 512GB of memory, and the M5 Ultra's 1.2 TB/s of memory bandwidth is fast enough to feed the 42 billion active parameters per token. Meta's line: two Mac Studios "behave like one supercomputer with a terabyte of unified memory." Mark Zuckerberg publicly thanked Apple CEO John Ternus for the collaboration.
Muse Spark Open: the desktop supercomputer math, checked
Let's do the arithmetic Meta skipped in the announcement. The 256GB Mac Studio already costs over $10,000, and the 512GB version Apple referenced hasn't even shipped yet — it's expected later this month. A two-machine setup with maxed-out memory will land somewhere in the mid five figures. "Run it on your desktop" is technically true and practically a flex for well-funded teams and labs, not something most developers will try at home.
That's fine, because the local angle isn't the point. The point is the API list. Muse Spark Open arrives day one on seven inference platforms, which means builders can start testing it against whatever they're using now without buying a rack of hardware. The local story is marketing with a kernel of truth: an 816B-parameter model that fits in a terabyte of unified memory is a genuinely new density class, and it puts pressure on the assumption that frontier-scale models need datacenter iron.
The open-weight race just got crowded
Muse Spark Open didn't land in a vacuum. It arrived in a 48-hour window that may go down as the busiest open-weight release stretch of the year. Liquid AI shipped LFM2.5-24B-A2B on October 13 — just 24 billion parameters total, only 2 billion active per token, and it reportedly beats Alibaba's Qwen3.6-35B on tool-use tasks. NVIDIA released the 120-billion-parameter Nemotron 3.5 Super VL multimodal model the same day as Meta. And Chinese lab Stepfun open-sourced Step 5 Preview, a 600-billion-parameter sparse MoE model that climbed to number two on OpenRouter's weekly usage chart with roughly 18.8 trillion tokens consumed in seven days, at roughly one-eighth the per-task cost of Claude Opus 5.
The pattern is worth naming: the open-weight market is fragmenting into specialists instead of chasing one parameter-count crown. Meta goes huge and general. Liquid AI goes tiny and efficient. NVIDIA goes multimodal and enterprise. Stepfun goes cheap and massively adopted. Benchmark tracker BenchLM counted 22 confirmed model launches in October alone, with 221 models tracked. When everyone's shipping, nobody gets to coast on a launch announcement.
What to actually do with Muse Spark Open
If you build with models, the practical move is boring: benchmark it on your actual workload this week. Run it on OpenRouter or Groq next to whatever you're using now — especially for agentic coding and long-context tasks, where the 1-million-token window and the coding-focused tuning are supposed to pay off. Don't rewrite your stack around it; open-weight releases are arriving weekly, and the winner of October is not guaranteed to be the winner of December.
Two things to watch. First, the license: "limited commercial use" is Meta's phrasing, and it's deliberately vaguer than a clean Apache 2.0 or MIT license. Read the actual terms before you ship anything that touches revenue. Second, the benchmark picture is still vendor-supplied. Meta says it excels at extended tasks and multi-step reasoning, but independent evaluations on SWE-bench-style coding tests will tell the real story over the next few weeks. The AI agent wars are being fought on exactly these claims, and the vendors always grade their own homework.
The bigger story is structural. A U.S. lab shipping its most capable open model yet, tuned for a consumer desktop, in the same week a Chinese open model hit number two on global usage charts — that's the open-weight race becoming a real market instead of a hobbyist sideshow. Retrieval-augmented pipelines and smaller specialist models like NVIDIA's Nemotron Bolt already gave builders options. Muse Spark Open adds a genuinely heavy option to the same menu. Choice is good. Just make sure you're choosing based on your benchmarks, not their press releases.
