Key points
- Meta releases Muse Spark Open's weights
- Where Meta's model ranks against China's
- How to run Muse Spark Open, including on Mac Studios
- Meta's license clause for big AI hosts
Meta Platforms (META) released the weights for Muse Spark Open on Thursday, calling it "the most capable US open-weight model to date." The model posts the highest score of any open-weight model built outside China on the Artificial Analysis Intelligence Index, though seven Chinese models score higher.
Muse Spark Open is an open-weight version of Meta's Muse Spark model, built for coding and agent work. Meta describes it as natively multimodal, meaning it was trained to handle images as well as text. It has 816 billion parameters, with 42 billion active at a time, and a 1-million-token context window. Developers can download the weights from Hugging Face and from cloud platforms run by Amazon (AMZN), Google, and Microsoft (MSFT). Hosts such as Groq, Nebius (NBIS), and Together AI also offer the model.
Seven Chinese models rank higher
As of October 15, Muse Spark Open scored 39 on the Artificial Analysis Intelligence Index, which combines 10 evaluations. It tied DeepSeek V4.1 Flash for eighth place on the firm's open-weight leaderboard.
Xiaomi's MiMo-V2.6-Pro led at 46, followed by Z AI's GLM-5.3 at 45. StepFun's Step 5 Preview and Moonshot AI's Kimi K3 each scored 44.
Artificial Analysis also described Muse Spark Open's hosted API pricing as expensive compared with open-weight models of similar size. It listed prices of $1.25 per million input tokens and $4.25 per million output tokens.
Muse Spark Open runs on two Mac Studios
Meta says Muse Spark Open is the first model in its class that can run on a desktop, using two Apple (AAPL) Mac Studios with M5 Ultra chips and Apple's MLX software. Each M5 Ultra's graphics processor can use up to 512GB of memory, so two machines can hold the entire model, according to Meta. The company says it tested agent tasks running entirely on two Mac Studios through its Muse Code tool, including bug fixes and new features for an open-source iOS app.
In a research post on Thursday, Apple said the model needs about 493GB for its parameters alone in a compressed 4-bit format. That's too much for one machine once the memory for computations is added. Apple measured speed on a cluster of four Mac Studios, where the model generated more than 36 tokens per second.
How to use Muse Spark Open
The simplest route is a hosted API. Meta lists Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI as launch providers, and Artificial Analysis lists the per-token prices above.
Developers who want to run the model themselves can download the weights from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, or Microsoft Foundry. Meta says it works with common inference software, including vLLM, SGLang, Nvidia's TensorRT-LLM, llama.cpp, and Apple's MLX, with optimizations for AMD, Apple, Intel, and Nvidia chips.
Running it locally takes serious hardware. At full precision, the parameters alone need more than 1.6TB of memory, Apple said, more than three times the 4-bit figure. Short of a multi-GPU server, the smallest setup Meta describes is two Mac Studios with M5 Ultra chips.
Big AI hosts need a separate deal with Meta
Meta released the weights under its own Muse Model License, which Meta says has "limited commercial terms, which allow many developers to use the model for free." Every product that uses the model must display "Built with Muse Spark Open" in its interface, documentation, or settings.
The license also covers companies that sell access to AI models through an API or a similar service. If such a company's total gross revenue, including its affiliates, tops $20 million over any 12 months, the company must notify Meta and sign a separate agreement before offering Muse Spark Open that way. Internal use, research, and products that use the model only to power their own features are exempt from that separate-agreement requirement. Meta says the model isn't meant for anyone under 18.
The release comes three days after Meta launched Muse Wonder and Muse Wonder Full, its more capable models, without publishing their weights. Meta Superintelligence Labs head Alexandr Wang said at the time, "We're still committed to open source."
Meta shares closed down 0.1% at $736.72 on Thursday, while Apple closed down 0.3% at $342.00.
