Meta has had a busy week. On Tuesday it launched Muse Wonder, its first frontier model, with the best version locked behind an API. On Thursday it did the opposite: it released Muse Spark Open, a model whose weights anyone can download, and called it “the most capable U.S. open-weight model to date.” Two launches, three days apart, and they tell you exactly where the AI industry is right now.
Muse Spark Open is a big model by any measure. It has 816 billion total parameters with 42 billion active per token, which makes it a mixture-of-experts design, and a context window of 1,048,576 tokens, or roughly a million. It is natively multimodal, accepting text, image, audio, and video inputs, and Meta built it for coding, code review, complex software engineering, multi-step reasoning, and long-running agentic tasks. MacRumors reported the launch on October 15, and Technobezz and TechBooky confirmed the details the same day.
Meta needed Apple to show it running on a desktop
The most unusual part of the announcement is not the model, it is the hardware demo. Meta worked with Apple to show Muse Spark Open running on a desktop setup: two 512GB Apple M5 Ultra Mac Studios, using the latest MLX framework. Apple’s unified memory architecture lets the M5 Ultra’s integrated GPU use up to 512GB of memory per machine, and Meta says the chip’s 1.2 TB/s of memory bandwidth is fast enough that two machines behave like one supercomputer with a terabyte of unified memory.
Meta CEO Mark Zuckerberg publicly thanked Apple CEO John Ternus for the collaboration, and Ternus replied that “Mac Studio keeps pushing the boundaries of on-device AI.” It is a remarkable moment: two companies that compete on headsets and messaging are teaming up to show that the biggest downloadable model of the year runs best on a Mac.
The catch: the license and the $20,000 desk
Before you plan your home AI lab, there are two catches worth knowing.
First, the license. Muse Spark Open ships under the Muse Model License, which Meta describes as having limited commercial terms. As TechBooky noted, a downloadable model is not the same as open-source software: you can download and run it for free, but developers should read the conditions before treating it as unrestricted. The weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.
Second, the hardware. The 512GB Mac Studio that Meta used for the demo has not even been released yet; it is expected later this month, and the existing 256GB Mac Studio already costs more than $10,000. So running the full model “on your own desktop” is real, but it is a desktop that costs over $20,000. For everyone else, hosted inference is the practical route: the model is live on Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, which was a day-zero launch partner.
Meta’s two-track week, in one table
The interesting move here is the pairing. Muse Wonder (the closed model) and Muse Spark Open (the downloadable one) launched within three days of each other. Here is how the week’s big models compare:
| Model | How you get it | Key specs | Price signal |
|---|---|---|---|
| Muse Spark Open | Downloadable weights + APIs | 816B total / 42B active, 1M context | Free to download; hosted pricing per provider |
| Muse Wonder | API only | Frontier model; best version locked away | $4 per million tokens (our breakdown) |
| Gemini 4 Argon | API only | General availability this week | $2/$10 per M tokens until Nov 30, then doubles (our breakdown) |
| Mistral Large 4 | Open weights | 1T total (“Le Chonk”) | Europe’s trillion-parameter bet (our breakdown) |
Early testers like the tool calling
Together AI, which made the model available on its serverless platform on launch day, shared early feedback from Factory and Ollama. Factory’s Zimu Li said the model “consistently selected the right tool and kept its responses compact,” while Ollama’s Dong Chen said it performed well on tool calling and long context, including multi-step and parallel calls with clean structured outputs.
Meta also shipped Helion GPU kernels that auto-tune themselves inside SGLang, with reported kernel-level speedups of 1.03x to 3.95x on NVIDIA and AMD GPUs, and a 3% end-to-end gain on AMD’s MI350 chip. And in Meta’s own testing, Muse Code paired with the model completed real iOS development work on Ice Cubes, an open-source Mastodon client, including fixes, features, and a complete build. It also operated DaVinci Resolve, examining individual video frames to pick highlights and produce vertical crops.
What this actually means
The open-weight race has quietly become the more interesting half of AI. Closed models keep getting more expensive to access at the top end, with rate cards and usage tiers multiplying. Meanwhile, Meta just put an 816-billion-parameter model in the public’s hands for free, tuned to run on consumer desktop hardware, on the same week it told API customers to pay for its best closed model.
That is not charity. Meta wants developers building on its weights instead of its rivals’, and it wants the inference ecosystem, from Together AI to Groq to Modal, competing to serve its model cheaply. For developers, the practical question is simpler: do you want the model you can inspect, host, and fine-tune yourself, or the slightly better one that lives behind someone else’s API? This week, Meta made you answer that question twice. Related: our look at Microsoft’s Maia 300 silicon bet, the other side of the same cost war.
