Muse Spark Open releases weights with substantial hardware needs

Meta released Muse Spark Open, a multimodal model with 816 billion total parameters, 42 billion active per token and a million-token context window. Its desktop demonstration uses two M5 Ultra Mac Studios with up to 512 GB each, not an ordinary laptop. Hosted providers offer another access route. The weights use Meta’s Muse Model License with limited commercial terms, rather than an unrestricted open-source license.

Meta AI Research · Oct 15, 2026

Why it matters

Open weights change deployment options, but memory and licensing still determine whether those options fit.

More context

Evaluate through a hosted provider before investing in local hardware. Meta lists MLX, SGLang and other inference engines; its demonstrated desktop workflow depends on a very large combined memory capacity.

Keep in mind

Provider costs vary; the launch page’s capability claims are vendor evaluations rather than independent comparisons.

Original read