Meta releases Muse Spark Open, an 816B open-weight AI for Macs

Meta released Muse Spark Open on October 15, 2026, calling it the most capable open-weight model available in the United States. At 816 billion parameters with a 1 million token context window and native multimodal support, it sits at the frontier of open AI — and Meta is offering it to anyone willing to run it locally.

The catch is the hardware requirement. Muse Spark Open needs at least 1 terabyte of unified memory to run, which means it performs best on Apple's Mac Studio with the M5 Ultra chip. In practice, that limits local deployment to well-funded teams and individuals with high-end workstations. API access through several partner platforms is available immediately for those who want the capability without the hardware investment.

Meta built Muse Spark Open specifically for complex coding and agentic tasks. It ships alongside Muse Code, Meta's AI-powered coding terminal, and the company demonstrated it fixing bugs in a Mastodon client and handling footage review and cropping in DaVinci Resolve with minimal human direction. The 1 million token context window means it can hold an entire large codebase in working memory during a session.

The release is a significant escalation in Meta's open-AI strategy. The company has long maintained Llama as an open-weight alternative to closed models from OpenAI and Anthropic, but Muse Spark Open is the first time Meta has released a frontier-scale model under open weights. By comparison, Meta's separately announced Muse Wonder is a closed API model targeting the same competitive tier.

Open-weight models carry real advantages beyond cost. Enterprises can audit the weights, run inference on-premises, and fine-tune without sending data to third-party servers. A frontier-scale open model narrows the gap between what is possible with a proprietary API and what is possible in a fully controlled environment — which matters for regulated industries and governments that have been reluctant to adopt closed AI.

Muse Spark Open is available now via API on multiple platforms. Local deployment guides for the M5 Ultra Mac Studio are published in Meta's documentation, as first reported by TechCrunch.