Meta Releases Muse Spark Open: An 816B Open-Weight AI Model

· Cortex Solutions

Two compact desktop workstations on an office desk beneath a glowing neural network, illustrating an open-weight AI model running locally

Meta released Muse Spark Open on Oct 15, 2026: an 816B open-weight model with a 1M-token context that runs on two Mac Studios. What it means for Saudi firms.

On Thursday, October 15, 2026, Meta announced "Muse Spark Open", an open-weight version of its Muse Spark model. In its research blog, the company called it "the most capable US open-weight model to date". The release was covered the same day by TechCrunch, among others, and by the Hugging Face blog, which published its technical details and how to run it.

"Open-weight" means companies and developers can download the model itself and run it on their own servers or machines, instead of only reaching it through a cloud service owned by someone else. That makes this news important for any organisation that cares where its data lives and what AI will cost it over the long run.

What exactly did Meta announce?

According to the Meta and Hugging Face blog posts published on October 15, 2026, these are the model's key specifications:

  • Around 816 billion parameters, but with a Mixture-of-Experts architecture, so only about 42 billion are active per token.
  • Each layer has 256 "experts", 8 of which are selected per token, according to Hugging Face.
  • A context window of up to one million tokens, meaning it can handle very long documents and files in one go.
  • It accepts text, images, audio and video, and outputs text.
  • It is built for complex coding and for "agentic" tasks in which the model carries out a sequence of steps, according to TechCrunch.

It runs on two desktop machines

Meta's main highlight is that the model can run locally on two Mac Studio machines with Apple's M5 Ultra chip, each with up to 512 GB of unified memory, using the MLX framework. Meta wrote that the two machines "behave like one supercomputer with a terabyte of unified memory". TechCrunch, citing Engadget, said it needs at least 1 TB of memory.

According to TechCrunch, Meta showed practical examples, including using the model with its Muse Code coding tool to fix an app, add features and do a full build, and using it with DaVinci Resolve to review and crop video footage.

On servers, Hugging Face says the full BF16 version is about 1.5 TB, and the compressed NVFP4 version can run on a single node of NVIDIA B200 GPUs. It also gives an example of fine-tuning the model on an organisation's own data with LoRA.

How can it be accessed?

According to Meta, the model is available in two ways:

  • Through APIs from providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
  • Or by downloading the weights from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.

Points worth noting

  • Licence: the model is released under the "Muse Model License", whose commercial terms Meta describes as limited. So "open-weight" here does not mean "open source" under a permissive licence, and the licence should be read before any commercial use.
  • Performance: Meta says the model beats popular open-weight models, but these are the company's own evaluations and have not yet been independently verified.
  • Competition: The New Stack described the launch as Meta's answer to China's best open-weight models, a sign of how fast this market is moving.
  • Cost: self-hosting is not free. The model needs a great deal of memory, and Hugging Face notes that the context cache alone can reach about 266 GB when the full one-million-token window is used.

What this means for Saudi businesses

  • A wider choice for data sovereignty: open-weight models make it possible to run AI on-premises or in a local cloud, so sensitive data never leaves your environment. This matters for sectors that handle personal or financial data and must comply with the Kingdom's Personal Data Protection Law.
  • Start by testing, not buying: before investing in expensive hardware, try the model through one of the API providers on real use cases from your business, especially its understanding of Arabic and local dialects, because the announcement gave no Arabic-specific results.
  • Review the licence with your legal team: limited commercial terms may affect how you can use the model inside a product you sell to customers.
  • Think about AI agents: the long context window and agentic capabilities open the door to automating tasks such as analysing long contracts, reviewing reports and supporting development teams.

At Cortex Solutions we help organisations choose the right model, in the cloud or on-premises, and build AI automation connected to their real business systems. Learn about our service: https://cortex.sa/services/ai-automation

Sources: Meta research blog (October 15, 2026), TechCrunch (October 15, 2026), Hugging Face blog (October 15, 2026), The New Stack.

Read next