News

Meta Releases Muse Spark Open Weights; Two Mac Studios Can Run It Locally (October 16, 2026)

On Oct 16, 2026, Meta released Muse Spark Open, an 816B-parameter MoE multimodal open-weight model (42B active, 1M tokens), optimized with Apple's MLX to run on two Mac Studios.

One-line summary

Meta announced on October 15, 2026 and released weights on October 16 for Muse Spark Open, an open-weight model: a native multimodal model with a mixture-of-experts architecture, 816 billion total parameters (42 billion active), and a 1 million token context window; optimized with Apple via the MLX framework, the full model can run on two Mac Studios with 512GB unified memory each.

Event Overview ​

According to Cailianshe's report of October 16, 2026, Meta released the open-weight model Muse Spark Open on Thursday (October 15) and partnered with Apple to run it locally on Mac computers. Meta AI Research's announcement calls the model "the most capable US open-weight model to date." Meta CEO Mark Zuckerberg posted on social media in the early hours of October 16 (Beijing time) that the weights were releasing that day, sharing a photo of the model running on two Mac Studios.

Key Information ​

ItemDetail
Official announcementOctober 15, 2026 (Meta AI Research)
Weights releasedOctober 16, 2026 (Zuckerberg: weights releasing that day)
ArchitectureMixture-of-experts, native multimodal
Parameters816 billion total, 42 billion active
Context window1 million tokens
FocusCoding and agentic tasks; long-horizon tasks, complex software engineering, multi-step reasoning
Claimed resultsMeta's own tests rank it first in 6 comparisons (workflows, tool use, knowledge reasoning, knowledge reliability, complex visual reasoning, visual factual QA)
Apple collaborationOptimized with MLX for Mac Studio with M5 Ultra; two Mac Studios with 512GB unified memory run the full model; the 512GB version is expected to go on sale in late October
Other supportOptimized for NVIDIA, AMD and Intel chips; compatible with multiple inference engines; available via API or self-hosted weights

Official Sources ​