News · SI
Meta has released the weights for Muse Spark Open, an 816B-parameter coding and agent model with a 1M-token context that runs locally on two Apple M5 Ultra Mac Studios. Here is what it means for teams deciding between hosted and self-run models.
Meta has released the weights for Muse Spark Open, which it calls the most capable US open-weight model to date[^1]. It is a natively multimodal model with 816 billion total parameters, 42 billion active, and a 1 million token context window, built for coding and agent work[^1].
01 / 04
The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted APIs are live at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI[^1][^2]. Meta says it runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, and it ships custom Helion GPU kernels in SGLang that Meta measured at 1.03 to 3.95 times faster at kernel level on NVIDIA and AMD GPUs[^1].
02 / 04
The headline trick is local deployment. Each Apple M5 Ultra Mac Studio can give its GPU up to 512 GB of unified memory, so two machines hold the whole model, and Meta says the pair behaves like one machine with a terabyte of memory[^1]. CNBC reports customers can link two to four Mac Studios, and that Apple published benchmarks and documentation on splitting open models across machines with MLX[^3].
Meta tested agent workflows running entirely on-device, including shipping real iOS changes to the open-source Mastodon client Ice Cubes and driving DaVinci Resolve to find and crop video highlights[^1].
03 / 04
Muse Spark Open is released under the Muse Model License, which Meta describes as having limited commercial terms that allow many developers to use it free[^1]. That is not the same as a permissive open-source licence, so check it before you build a paid product on top.
Meta also says the model had extra safety post-training and met the release thresholds in its Meta Superintelligence Scaling Framework, with a full evaluation report still to come[^1].
04 / 04
If your team has been blocked from frontier-grade coding agents because code or customer data cannot leave your network, this is a real option: a model in this class on hardware that sits on a desk. The trade-off is cost and setup. Two top-spec Mac Studios are a serious purchase, and hosted endpoints from the listed providers will still be the faster way to evaluate it. Start with a hosted trial on your own repos, then decide whether self-hosting pays for itself.
FAQ
Muse Spark Open is the open-weight version of Meta's Muse Spark model, released on 15 October 2026 (US time). It is natively multimodal, has 816 billion total parameters with 42 billion active per token, and supports a 1 million token context window. Meta built it for coding, long-horizon agent work and multi-step reasoning.
Yes, but not on one. Meta says it runs on two Apple M5 Ultra Mac Studios using Apple's MLX framework, because each Mac Studio's GPU can use up to 512 GB of unified memory. CNBC reports customers can pair between two and four Mac Studios.
Hosted access is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
Not without limits. Meta is releasing it under the Muse Model License, which it describes as having limited commercial terms that let many developers use the model for free. Teams should read the licence before building a paid product on it.
Meta lists SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimisations for AMD, Apple, Intel and NVIDIA hardware. It also ships custom Helion GPU kernels in SGLang.
Meta says the model received additional safety post-training because open weights carry extra risk, and that it met the release thresholds in its Meta Superintelligence Scaling Framework. Detailed results are in the model card, with a full evaluation report still to come.
That’s the story
Spectrum Web Co ships SI and Web3 products, places talent and builds social narratives as monthly services.
