Meta releases Muse Spark Open weights, an 816-billion-parameter model that runs on two Mac Studios
Meta, Thursday: Muse Spark Open, 816 billion parameters in total and 42 billion active, a 1 million-token window, built for coding and agents. It runs on two Apple M5 Ultra Mac Studios with MLX. Weights on Hugging Face, SageMaker, Model Garden and Foundry. Ina Fried, Axios: Apple helped Meta run it on two or four Mac Studios. Alexandr Wang says Daniel Gross helped bring the companies together.
Deep dive4 minIn-depth
James Okonkwo4 min read
- Meta
- Muse Spark
- Open weights
- Apple

Photo: Meta AI ResearchSource page
Meta on Thursday released the weights for Muse Spark Open, an open-weight model it says is built for coding and agentic work. In a research post the company called it a natively multimodal system with 816 billion parameters in total, 42 billion active at a time, and a context window of 1 million tokens. It called the release the most capable US open-weight model to date, and the first in its class to run on a desktop: two Apple M5 Ultra Mac Studios, using Apple’s MLX framework. Reuters, in a story timestamped 5:19 p.m. on Thursday and carried by Investing.com, reported the same parameter counts and the same context window, and said the release expands developers’ ability to download and adapt the technology.
The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted inference is live at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta said the model runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA chips. It also said custom Helion GPU kernels in SGLang deliver 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs, and a 3 percent end-to-end speedup on AMD’s MI350. Those speed figures are Meta’s. The license is the Muse Model License. Meta said it carries limited commercial terms and lets many developers use the model for free. Ina Fried at Axios wrote that the download is free, with a license that allows commercial use, and that Meta describes Muse Spark Open as similar to Muse Spark 1.3. Reuters noted that Meta released a smaller open-weight model, Muse Glimmer, in August, and that chief executive Mark Zuckerberg has argued advanced AI should not be concentrated among a small number of companies.
What the file actually is
Pedro Cuenca, in a Hugging Face post published Thursday, described the checkpoint in more detail. The model takes text, images, audio and video. It generates text. It is a mixture of experts: 256 experts in each mixture-of-experts layer, with a router selecting 8 for each token. That is how the active compute stays near 42 billion parameters while the full network stays near 816 billion, and why the weights still have to sit in memory at the larger size. Hugging Face said the BF16 checkpoint is about 1.5 terabytes in safetensors. A 4-bit NVFP4 checkpoint is aimed at NVIDIA Blackwell GPUs. In Transformers, Cuenca wrote, the BF16 weights run across four nodes of eight H200 GPUs, and the NVFP4 weights run on a single B200 node. Reasoning cannot be switched off. The effort setting is low, medium, high or “xhigh.”
Apple’s part of the release is the local run. Meta said the M5 Ultra’s integrated GPU can use up to 512 GB of unified memory on one Mac Studio, so two machines can hold the model, and that the chip’s 1.2 terabytes per second of memory bandwidth is enough to read the slice of weights each token needs. Two Mac Studios, Meta wrote, then behave like one machine with a terabyte of unified memory. Meta said it tested agentic work running entirely on the machines, including shipping iOS changes — bug fixes, new features and a full build — to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically. Axios reported a wider rig: Meta worked with Apple so the model runs locally across two or four connected M5 Ultra Mac Studios, again on MLX.
Alexandr Wang, Meta’s AI chief, told Fried the companies have been working together for several months, and that Daniel Gross, a former Apple executive now at Meta, helped bring them to the table. “I think he played a really big role in kind of getting the two companies to work nicely together,” Wang said in a telephone interview. He also told Axios: “We’re excited for mommy and daddy to be playing nice again.” Because the model runs locally, he said, “Meta has no control or visibility to any of that,” meaning developers’ prompts and use. He told Axios the companies will keep competing in some areas, and that “where it makes sense to collaborate and partner, we want to be able to do so.” Fried’s account recalls the public clashes between Apple executive chairman Tim Cook and Zuckerberg over privacy, including Cook’s 2018 line, amid the Cambridge Analytica scandal, that if he were Zuckerberg he “wouldn’t be in this situation,” which Zuckerberg called “extremely glib.”
Meta said open weights need their own safeguards. Muse Spark Open had additional safety post-training, the research post said, and the company evaluated it under the Meta Superintelligence Scaling Framework, covering what it called the most severe risks. Meta said the model met the thresholds that framework requires for a release. Detailed results are in the model card. A fuller evaluation report, Meta said, is still to come. Axios wrote that publishing weights of this size raises the concern that someone could remove safeguards and use the system for cyberattacks or biological threats. Meta’s post does not say the weights cannot be changed. It says the release is how it wants models of this class to be runnable at home.


