Meta says it released the weights for Muse Spark Open on October 15, calling it "the most capable US open-weight model to date." Meta lists 816B total parameters, 42B active, a 1M-token context window and native multimodality. The weights are downloadable from Hugging Face and three cloud platforms, and seven API hosts offer it.

What it runs on

Meta says it optimized the model for Mac Studio with M5 Ultra using Apple's latest MLX framework. It says the chip can use up to 512 GB per machine, so two Mac Studios can hold the whole model and behave like one machine with a terabyte of unified memory.

Apple's Machine Learning Research post gives the memory math. In its native BF16 precision, Apple says, the model needs more than 1.6TB for parameters alone. In 4-bit NVFP4 it needs about 493GB. With working memory added, Apple says it cannot fit on one machine and needs two or more Mac Studios.

Apple reports that four M5 Ultra Mac Studios generate over 36 tokens per second at an 8,192-token context. Its table lists 15.74 tokens per second on two machines using pipeline parallelization, 24.99 on two using tensor parallelization and 36.73 on four. Apple says the 4-bit NVFP4 version is what it tested.

How it compares with other local models

Apple's post also tests two models that fit on a single 512GB Mac Studio. Qwen 3.8 (27B) and Mistral Medium 3.5 (128B) take up to 68GB and 158GB at a 65,536-token context, Apple says. At 8,192 tokens on one M5 Ultra, its table lists 19.69 tokens per second for Qwen and 7.70 for Mistral.

Apple describes Muse Spark Open as a mixture-of-experts model that activates 42B parameters per token, while Qwen and Mistral are dense. As of Oct. 15, Apple's published comparison covers only Mac Studio configurations, and it does not test Muse Spark Open on Macs with less memory.

Price and availability

Engadget reports that, as of Oct. 15, Apple has not announced a price or ship date for the 512GB Mac Studio, saying only "late October," and that four 256GB M5 Ultra machines would cost roughly $38,000, Engadget's estimate. It adds that Meta's own benchmarks put the model ahead of GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning.

Meta says the weights ship under the Muse Model License, with "limited commercial terms" that let many developers use the model for free. 9to5Mac also reports the Mac Studio optimization and links Apple's MLX post.

Why it matters

Our assessment: "Open weights on your desktop" here means a desk-sized cluster. Apple tested two 512GB Mac Studios, with four machines delivering higher throughput. Apple's post gives no result for smaller Macs as of Oct. 15.

The benchmark comparison with GLM 5.3 and Kimi K3 is, per Engadget, Meta's own. As of Oct. 15, Meta's post and Apple's post report no independent test of Muse Spark Open on a Mac.