Skip to content
Neural Daily
hu

ModelsNews

Meta's Muse Spark Open: open weights that run on Macs

Meta released the weights for Muse Spark Open on October 15: a natively multimodal model with 816 billion total parameters (42 billion active) and a 1-million-token context window. Meta calls it the most capable US open-weight model to date, a claim based on its own evaluations.

By: Mate · · Updated: · 3 min read

Blue flowing lines on a pale background with the headline Muse Spark Open Weights
Image: Image: Meta (https://research.meta.ai/blog/muse-spark-open)

Meta published the weights for Muse Spark Open on Thursday (announcement). It is aimed at coding and agentic work: long-horizon tasks, complex software engineering and multi-step reasoning.

The headline specs: 816B total parameters with 42B active per token (a mixture-of-experts design), a 1M-token context window, and native multimodality.

Two Mac Studios, one model

The hardware angle is the most interesting part. Meta says it is the first model in its class to run on a desktop: two Apple M5 Ultra Mac Studios using the MLX framework. Each machine's GPU can address up to 512 GB of unified memory, so together they offer about 1 TB, enough to hold the whole model.

Why it matters: this is not "runs on a laptop". TechCrunch, citing Engadget, puts the requirement at at least 1 TB of RAM (TechCrunch). But for a small team or a university lab, that is a desk, not a rented GPU cluster. Where sensitive data can't go to the cloud, that is a real difference.

Where to get it

Hosted APIs are available from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Weights can be downloaded from Hugging Face (collection), Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry.

Supported inference engines include SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth. Meta says its Helion GPU kernels for SGLang deliver 1.03–3.95× kernel-level speedups on NVIDIA and AMD GPUs, but only a 3% end-to-end gain on AMD MI350. Both are Meta's own figures.

Demos and caveats

Meta showed two on-device demos using its Muse Code terminal: shipping real iOS changes to Ice Cubes, an open-source Mastodon client (bug fixes, new features, full build), and driving DaVinci Resolve end to end, with the model reviewing footage frame by frame and cropping vertical clips.

What's missing: the "most capable US open-weight model" label rests on Meta's own benchmarks, with no independent evaluation yet. A full safety evaluation report is promised "soon"; until then the model card has the details.

Meta benchmark scorecard comparing Muse Spark Open with other open-weight models on agentic, reasoning, coding and multimodal benchmarks
Image: Meta, vendor-reported (https://research.meta.ai/blog/muse-spark-open)

Source: Meta, vendor-reported (Meta AI Research)

The license matters too. Weights ship under the Muse Model License with limited commercial terms. This is not Apache 2.0, so have legal look at it before commercial use.

What to do now

If you have GPU capacity or two high-end Macs, it is worth benchmarking against a closed model on your own coding and agent tasks. Wait for independent evaluations and a license review before anything goes to production.

    Reviewed and edited by the Neural Daily desk. About

    Related articles