Meta released the weights of Muse Spark Open on 15 October 2026, an 816B-parameter model with 42B active parameters and a 1M-token context window, which Meta calls “the most capable US open-weight model to date.”
Meta says its licence allows many developers to use the model for free, but the licence sets a line: a company that offers model access through an API and has more than $20 million in gross revenue in any 12-month period must sign a separate agreement with Meta first.
Business Pill · QUANTIZATION
A one-minute explainer of quantization: how a large AI model is shrunk to fit in less memory. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, the licence is where the release meets Meta’s business: the weights are free for many developers, Meta says, but large companies selling model access by API must sign a separate agreement with Meta.
According to Meta’s research blog, Muse Spark Open is “a natively multimodal model with 816B total parameters (42B active) and a 1M-token context window.” Meta says it is built for coding and agentic work.
The model card describes it as a mixture-of-experts transformer with a perception encoder. It takes text, image, video and audio input and produces text, and it was trained on data from more than 100 languages, the card says.
Meta’s developer blog calls it “the open weights version of Muse Spark” and says it activates roughly 42 billion parameters per token.
The card says this open-weights version has additional safety post-training, and Meta recommends deploying it as part of a broader AI system with guardrails.
Under the Muse Model License, a company offering model access through an API whose gross revenue exceeds $20 million in any 12-month period must notify Meta and sign a separate agreement; Meta says Muse Spark Open has 816B total and 42B active parameters and a 1M-token context (Meta, 15 October 2026).
What Was Posted on X
Today weโre releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Meta AI’s launch post on 15 October 2026 calls Muse Spark Open ‘the most capable US open-weight model to date’ and lists 816B total / 42B active parameters, a 1M-token context window and native multimodality across text, image, video and audio. ‘Most capable’ is Meta’s own claim.
Mark Zuckerberg posted on 15 October 2026 that Muse Spark Open, ‘weights releasing today’, ‘runs great on new Mac Studios thanks to work with @johnternus and Apple’. The post names Apple’s John Ternus but gives no performance figures.
Congrats to the team! Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!
Apple CEO John Ternus replied to Mark Zuckerberg’s post on 15 October 2026: “Amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!” It is Apple’s chief executive publicly endorsing the Mac Studio as hardware for Meta’s open model; neither company has described any commercial terms.
The Licence Terms
Meta’s research blog says the weights are released “under the Muse Model License with limited commercial terms, which allow many developers to use the model for free.”
The licence text sets out the main limit. If a licensee provides third parties with access to AI models through an API, and its aggregate gross revenue “exceeds twenty (20) million US dollars” in any 12-month period, it “must notify Meta and enter into a separate written agreement with Meta” before offering the model as part of that business.
The licence says that clause does not apply to internal use, research, commercial activity that is not model-as-a-service, or using the model to power features of the licensee’s own product.
It also requires products built on the model to display “Built with Muse Spark Open”, and it says “Meta claims no ownership of outputs generated by the Software.”
Running It on Two Mac Studios
Meta says Muse Spark Open is “the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework.”
The research blog says each M5 Ultra Mac Studio can give its GPU up to 512 GB of unified memory, so two machines hold the whole model, and that “two Mac Studios behave like one supercomputer with a terabyte of unified memory.”
The developer blog says MLX runs the NVFP4 checkpoint on two Mac Studios with 512GB each, or four with 256GB each, for a total of 1TB.
On GPUs, the model card says the model needs at least 700GB of VRAM for NVFP4 and 2TB for BF16.
Where It Is Available
Meta lists hosted inference at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
It says the weights are available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models.
The developer blog says the model runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth at launch, verified on NVIDIA, AMD, Intel and Apple silicon.
Meta also says the release ships Helion GPU kernels in SGLang with 1.03-3.95x kernel-level speedups on NVIDIA and AMD GPUs.
What Meta’s Benchmarks Show
Meta’s model card compares Muse Spark Open with other open-weight models, including Kimi K3, GLM 5.3, Qwen3.8-2.4T-A95B and Nemotron 3 Ultra.
On Meta’s table, Muse Spark Open scores 61.8 on JobBench and 89.2 on MCP-Atlas, the highest in those rows. On the Artificial Analysis Intelligence Index it scores 39, against 45 for GLM 5.3 and 44 for Kimi K3.
On Terminal-Bench 2.1 the card lists 76.4 for Muse Spark Open and 85.0 for Kimi K3. These are Meta’s own figures.
As we read it, the $20 million clause draws the line at resale: the licence exempts internal use, research and powering a licensee’s own product, and applies to companies offering the model to others as a service.
As we read it, the distribution list is wide: Meta names seven hosted inference providers and lists the weights on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models, the catalogues of three large cloud providers.
As we read it, the Mac Studio target is a hardware statement: Meta says two M5 Ultra machines hold the whole model in a terabyte of unified memory, against a GPU minimum of 700GB of VRAM for NVFP4 in the model card.
Meta also sells paid Muse Wonder plans from $20 to $200 a month, as we reported on 12 October.
Meta, in its 15 October 2026 research blog
“We believe the future is for everyone, and that includes the freedom to run advanced models wherever, including at home.”
Three Implications
MODEL-AS-A-SERVICE HOSTS The licence says hosts above $20 million in gross revenue over 12 months must notify Meta and sign a separate agreement before offering the model.
DEVELOPERS BUILDING PRODUCTS The licence requires products built on the model to display ‘Built with Muse Spark Open’, and says Meta claims no ownership of outputs.
OPEN-WEIGHT RIVALS Meta’s model card compares Muse Spark Open with Kimi K3, GLM 5.3, Qwen3.8-2.4T-A95B and Nemotron 3 Ultra, ahead on some rows and behind on others.
The framework sets out the pattern as: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize).”
As we read it, Meta’s release splits the two: the weights are open, while the licence keeps a separate agreement for large model-as-a-service businesses, and Meta also sells paid Muse Wonder plans. The framework does not discuss this release.
Meta’s model card gives GPU VRAM minimums of 2TB for BF16 and 700GB for NVFP4; Meta’s research blog says two M5 Ultra Mac Studios, with up to 512 GB each, hold the whole model in a terabyte of unified memory (15 October 2026).
What Is Not Established
The texts we read give no pricing for the separate agreement that large API hosts must sign, and no list of companies that have signed one.
The benchmark results are Meta’s own, and the full evaluation report is described as available soon. Meta’s ranking claim is its own.
Business Engineer Framework
The Open vs Closed Meta-Framework
A Business Engineer framework on which layer of a stack to keep closed and which to open.
Meta has released the weights of Muse Spark Open, an 816B-parameter model with 42B active parameters and a 1M-token context, which it says runs on two Mac Studios.
Meta says the Muse Model License allows many developers to use the model for free; under the licence, a company offering model access by API with more than $20 million in gross revenue in any consecutive 12-month period must sign a separate agreement with Meta.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. We read Meta’s 15 October 2026 research blog post “Introducing Muse Spark Open”, its developer blog “Build with Muse Spark Open”, and the model card and Muse Model License in full. Meta’s own Hugging Face repository is gated, so we read the model card and licence in the copies reproduced in Unsloth’s mirror repository. The benchmark figures and the ranking claim are Meta’s own. The Nemotron, MAI-Base-2 and Muse Wonder context comes from our own earlier reporting. We did not use other outlets’ reports.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
