Meta's open-weight Muse Spark has 816B parameters. Two Mac Studios can run it

What is actually released

Meta's post calls Muse Spark Open "the most capable US open-weight model to date". It is natively multimodal with 816B total parameters, 42B active per token and a 1M-token context window, aimed at coding, agentic work and multi-step reasoning. The Muse Model License has limited commercial terms that, Meta says, let many developers use it free of charge.

Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted APIs come from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and the model works with SGLang, vLLM, TensorRT-LLM, llama.cpp and Unsloth. Custom Helion GPU kernels in SGLang give 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs, with a 3% end-to-end gain on AMD MI350.

The local-hardware claim and what is missing

Two Mac Studios give about 1 TB of combined memory, enough for the weights. Meta demonstrated this on iOS changes to the Ice Cubes app and video editing in DaVinci Resolve, which are demos on specific tasks, not general performance measures. The page's "Key Benchmarks" section carries no figures in its text and links to an evaluation report that is not yet published.

The one outside score puts it level with smaller models

With no published report, the thread supplied its own comparison. A former lead of Google's gRPC team read the third-party Artificial Analysis ranking:

"Artificial analysis shows this: underperforming Qwen3.8 Flash Next and GLM 5.3 Flash, both of which you can run on a single B300 ." — former lead of Google's gRPC team

A commenter who transcribed the leaderboard pasted the figures from that ranking: Muse Spark Open at 39, the same as DeepSeek V4.1 Flash and below Meta's hosted Muse Spark 1.3 at 45. These are a reader's transcription of a screenshot, not Meta's numbers. On the license, a commenter who read the text and quoted the clause found that it bites only providers of model-as-a-service above $20 million in gross revenue:

"They're allowing the use of this in commercial products, just not selling the model directly." — commenter who read and quoted the license clause

For a developer, the license is probably not the obstacle; the open question is whether 816B parameters on a terabyte of memory beats a model that fits one GPU. The only independent evidence so far says it does not, and until Meta's own report lands, that is the number to beat.