Meta open-sources Muse Spark Open: 816B total / 42B active parameters and a 1M-token context, with Meta saying it runs locally on two M5 Ultra Mac Studios

Start with what the official announcement actually says. In its October 15 release post, Meta AI Research describes Muse Spark Open as the most capable US open-weight model to date: a natively multimodal model with 816B total parameters (42B active) and a 1M-token context window, built for coding and agentic work and said to excel at long-horizon tasks, complex software engineering and multi-step reasoning. The official model card on Hugging Face is more specific: release date October 15, 2026; 816B total / 42B active parameters; a mixture-of-experts transformer with a perception encoder; a 1M context length at inference; a vocabulary size of 202,048; and modalities listed as text, image, video and audio input with text output. In other words, it can read screenshots, recordings and footage but writes text, and it is not an image, audio or video generator.

The second concrete point is where to get it. The official post says Muse Spark Open is available today through API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, while model weights can be downloaded on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models. On the inference side, Meta says it runs on engines including SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimisations for AMD, Apple, Intel and NVIDIA silicon, and that developers can further customise it with PyTorch. On licensing, Meta says the weights ship under the Muse Model License with limited commercial terms, which it describes as allowing many developers to use the model for free. The licence text on Hugging Face states that if a licensee offering a Model-as-a-Service business and its affiliates exceed 20 million US dollars in aggregate gross revenue over any consecutive twelve-month period, it must notify Meta and enter a separate written agreement before offering the software that way; internal use, research and commercial activity that is not Model as a Service fall outside that clause.

The third point is the Apple silicon angle, and it is the most counterintuitive part of the release. Meta says Muse Spark Open is the first model in its class to run on a desktop, using two M5 Ultra Mac Studios with the latest MLX framework. The mechanism is spelled out: Apple's unified memory architecture lets the M5 Ultra's integrated GPU use up to 512GB per Mac Studio, so two machines can hold an entire model, and the chip's 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs. Meta's framing is that two Mac Studios behave like one supercomputer with a terabyte of unified memory. To show this is not just a paper demo, Meta says it tested a range of agentic workflows running entirely on-device with Muse Code powered by Muse Spark Open, including shipping real iOS changes to Ice Cubes, an open-source Mastodon client, covering bug fixes, new features and a full build, and driving DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.

The boundaries deserve the same clarity. First, this does not simply flip the open-weight pecking order. VentureBeat reports that the model card lists an Artificial Analysis Intelligence Index score of 39 for Muse Spark Open, below the independently published 46 for MiMo-V2.6-Pro; and on Meta's own table, Spark Open leads the listed rivals on JobBench and MCP-Atlas tool-use tests but trails Kimi K3 and GLM 5.3 on terminal coding, extended software-engineering tasks and the GDPVal-AA professional-work evaluation. Second, the desktop path still has real friction: VentureBeat notes Apple's 512GB memory configuration does not arrive until late October and its price is unannounced, while the model card lists at least 700GB of accelerator memory for the NVFP4 version and 2TB for BF16, so activating a smaller share of the model does not shrink memory requirements to those of a standalone 42B model. Third, the blog gives no hosted token pricing and no end-to-end price for the two-machine setup. Fourth, on safety, Meta says the model was evaluated under its Meta Superintelligence Scaling Framework, covering the most severe risks, and met the thresholds required for release, with detailed results in the model card and a full evaluation report to follow.