Meta Muse Spark Open Review (2026)

Updated: October 2026 · 7 min read · By the CreatorSprout team

Image: Meta Platforms, Inc. (Wikimedia Commons, public domain)

Launch coverage based on Meta's release materials and press reporting — we have not tested this model hands-on. Specs and pricing verified October 2026; confirm on official pages before you buy. No affiliate links in this article.

The headline for creators and builders: this is the first model in its class that Meta says can run on a desktop — two Mac Studios, to be exact. Here's what the launch actually contains.

What Muse Spark Open is

Muse Spark Open is an open-weight version of Meta's Muse Spark model family, designed specifically for coding and agentic workloads — the kind of multi-step work where a model plans, calls tools, and completes extended tasks. Meta describes it as natively multimodal, trained to handle images as well as text.

The release matters because of timing: it arrived October 15, three days after Meta launched Muse Wonder and Muse Wonder Full — its more capable models — without publishing their weights. Releasing a huge open-weight model in the same week is Meta signaling it is, in Alexandr Wang's words, "still committed to open source."

The specs

816 billion total parameters — 42 billion active per token (mixture-of-experts design, so only a slice is computed for each token)
1-million-token context window — room for entire codebases or long documents in one pass
Natively multimodal — trained on images as well as text, per Meta
Built for coding and agents — Meta says it excels at extended tasks, multi-step reasoning, and software engineering
Framework support — works with vLLM, SGLang, NVIDIA's TensorRT-LLM, llama.cpp, and Apple's MLX, with optimizations for AMD, Apple, Intel, and NVIDIA chips (per Meta's release materials)
Weights — downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models; hosted APIs via Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI

The Mac angle: runs on two Mac Studios

This is the most unusual part of the launch. Meta worked directly with Apple to tune Muse Spark Open for the Mac, and says two 512GB Apple M5 Ultra Mac Studios running the MLX framework can hold the entire model — "the best way to run a powerful model on your own desktop," in Meta's words.

The numbers: Apple says the model's parameters need about 493GB in a compressed 4-bit format — too much for one machine once compute memory is added, but two 512GB Mac Studios hold it. Apple measured throughput on a four-Mac-Studio cluster at more than 36 tokens per second. At full precision, the parameters alone would need over 1.6TB of memory.

Meta even tested real agent work on the setup, using its Muse Code tool to fix bugs and add features to an open-source iOS app entirely on two Mac Studios. Caveats: the 512GB Mac Studio hadn't shipped yet at launch (expected later in October), and the existing 256GB model is already priced over $10,000. This is not a budget setup.

Pricing and how to use it

Three ways in:

Hosted API (easiest). Launch providers include Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Artificial Analysis listed prices of $1.25 per million input tokens and $4.25 per million output tokens — described as expensive for an open-weight model of this size.
Download the weights (free). Available on Hugging Face and the major cloud model gardens, free to download under Meta's license (see fine print below).
Run locally (serious hardware). Two M5 Ultra Mac Studios at minimum, per Meta — or a multi-GPU server.

Pricing was verified October 2026 via independent reporting; providers change rates often, so check before you commit.

The license fine print

Muse Spark Open is not Apache or MIT licensed. It ships under Meta's own Muse Model License, which Meta describes as having "limited commercial terms, which allow many developers to use the model for free." Two clauses worth knowing:

The $20M API-host clause. Companies that sell access to AI models through an API — and gross over $20 million in 12 months including affiliates — must notify Meta and sign a separate agreement before offering this model. Internal use, research, and products that just use the model to power their own features are exempt.
"Built with Muse Spark Open." Every product using the model must display that attribution in its interface, documentation, or settings.

Meta also states the model isn't meant for anyone under 18. For solo developers and small teams building their own products, the license is effectively free; if you're building a model-hosting business at scale, read the full license first.

How it stacks up

On the independent Artificial Analysis Intelligence Index (10 evaluations combined), Muse Spark Open scored 39 as of October 15 — the highest of any open-weight model built outside China, tied with DeepSeek V4.1 Flash. But seven Chinese models scored higher: Xiaomi's MiMo-V2.6-Pro led at 46, followed by Z.AI's GLM-5.3 at 45, with StepFun's Step 5 Preview and Moonshot's Kimi K3 at 44.

Context matters: the same week saw a pileup of open-weight releases — Liquid AI's LFM2.5-24B-A2B, NVIDIA's Nemotron 3.5 Super VL 120B, StepFun's Step 5 Preview weights, and six Jais 3 variants. "Most capable U.S. open-weight model" is accurate per the rankings — and also a reminder that the U.S. qualifier is doing real work in that sentence.

What it means for creators

Let's be honest about who this is for: developers building AI agents and coding tools, not the average creator editing videos. But it matters to creators indirectly — models like this are what power the agent features, code assistants, and automation tools you'll be buying next year. Open weights mean more competition, more fine-tuned variants, and eventually better tools at lower prices.

You should care if you: build software, fine-tune models, or run AI workloads on Apple hardware — the Mac Studio angle is genuinely new territory.

You can safely ignore it if you: just need a chatbot, writing assistant, or image generator — the hosted APIs for everyday models are cheaper and simpler.

Verdict

Muse Spark Open is the most interesting open-weight release of October 2026 — not because it's the strongest model on the planet, but because of what it enables: frontier-adjacent agentic capability that developers can actually download, and a credible path to running it on a desk. The $1.25/$4.25 hosted pricing is steep for open weights, and the license has real strings attached for model-hosting businesses. But for the open-source AI ecosystem, Meta releasing 816B parameters with honest licensing terms three days after keeping its flagship closed is the right kind of compromise.

FAQ

What is Meta's Muse Spark Open?

An open-weight AI model Meta released on October 15, 2026, for coding and agentic work. It has 816 billion parameters (42 billion active per token) and a 1-million-token context window. Meta calls it the most capable U.S. open-weight model to date.

Can Muse Spark Open run on a Mac?

Yes — on two machines. Meta says it runs on two 512GB Apple M5 Ultra Mac Studios using Apple's MLX software. The model needs about 493GB for its parameters in a 4-bit format; four Mac Studios generated over 36 tokens per second in Apple's testing.

Is Muse Spark Open free?

The weights are free to download under Meta's Muse Model License. Commercial use is allowed for most developers, but API-hosting companies with over $20 million in annual revenue must sign a separate agreement with Meta, and all products must display "Built with Muse Spark Open."

How does Muse Spark Open compare to Chinese open-weight models?

It scores 39 on the Artificial Analysis Intelligence Index — the highest for any non-Chinese open-weight model — but seven Chinese models score higher, led by Xiaomi's MiMo-V2.6-Pro at 46.

How much does the Muse Spark Open API cost?

Third-party hosts listed approximately $1.25 per million input tokens and $4.25 per million output tokens at launch (October 2026). Prices change; verify on the provider's page.