Meta Launches Muse Spark Open: 816-Billion-Parameter AI Model for Local Mac Running
Meta released Muse Spark Open on Wednesday (October 15), an artificial intelligence model with 816 billion parameters — of which 42 billion are activated per input — designed to run locally on Mac Studios equipped with the M5 Ultra chip. With a context window of up to 1 million tokens and native multimodal capabilities (text and image), the model positions itself, according to the company, as the most capable open-weight model produced in the United States to date.
The launch represents an unusual moment of cooperation between two companies that spend most of their time in a cold war of technological rivalry. Apple engineers worked for months with Meta's AI team to optimize Muse Spark Open for the MLX framework, which was built specifically to leverage the unified memory architecture of Apple Silicon chips. Apple CEO John Ternus himself highlighted the collaboration on social media, while Mark Zuckerberg posted that the model "runs great on new Mac Studios thanks to work with Apple."
Why 1 terabyte of RAM changes the game
The most striking number here is not the parameter count, but the memory required to run the model locally. At BF16 precision, Muse Spark Open requires more than 1.6 terabytes of memory — three times what the largest Mac Studio offers. Meta addressed this with 4-bit NVFP4 quantization, reducing the weights to about 493 GB. This means two Mac Studios with 512 GB can hold the entire model.
Two Mac Studios process the model at 24.99 tokens per second; four machines reach 36.73 tok/s. In practical terms, this is interactive-speed performance for coding and multi-step reasoning tasks.
The required infrastructure is expensive: Mac Studios with M5 Ultra and 256 GB start at $10,799. A four-machine setup reaches $38,000. The 512 GB Mac Studio — the recommended configuration — does not yet have an announced release date, only "late October." Even so, the cost per token over time could offset thousands of dollars monthly spent on proprietary model APIs for teams that burn through API budgets.
Benchmarks: Meta vs. China
Meta published benchmark results comparing Muse Spark Open with two leading Chinese models: Zhipu AI's GLM 5.3 and Moonshot AI's Kimi K3. On agentic and visual reasoning benchmarks, Muse Spark Open shows competitive or superior performance.
Standout score: 76.4 on Terminal-bench 2.1 — a benchmark focused on real-world software engineering tasks in the terminal. This positions the model among the best open-weight models available, competing directly with Llama and Qwen.
Real-world use cases: from code to video
Meta demonstrated Muse Spark Open in two concrete scenarios:
1. Software engineering: using Muse Code (Meta's coding terminal), the model applied bug fixes, added new features, and completed a full build of Ice Cubes — an open-source Mastodon client. All executed locally.
2. Video editing: the model drove DaVinci Resolve end-to-end, analyzing video frames frame by frame to identify the best moments and crop them vertically — a task that typically would require significant human intervention.
These examples matter because they go beyond the chatbot. Muse Spark Open is designed for agentic labor — autonomous work in complex multi-step workflows — not just answering questions.
The bigger picture: the open-weight war and AI sovereignty
The launch of Muse Spark Open does not happen in a vacuum. Meta has been consistent in its open-weight strategy since LLaMA. In recent months, the ecosystem has accelerated: Qwen 3.8-27B became popular for 24 GB Macs, Mistral Large 4 offers 1.05 trillion parameters, and Anthropic's Claude Haiku 5.5 arrived as a compact version at dramatically reduced costs.
What makes Muse Spark Open different is not the size — there are larger models. It is the optimization for Apple Silicon and the promise of running locally on your desk. The idea of personal superintelligence, popularized by Sam Altman, takes concrete form when you can run the most capable US model on physical hardware you control.
Meta frames the model as part of a larger narrative: the US should not let Chinese laboratories lead in open-source AI. But Meta's own open-weight strategy is, ironically, precisely what gives Chinese labs data to improve their models. The open-weight AI war is, ultimately, an inference data war — who runs more, learns more.
What to watch
Muse Spark Open is available for free download on Hugging Face under the Muse Model License (with limited commercial terms). It is also available via API on Together AI, Fireworks AI, Groq, Baseten, Modal, Nebius, and OpenRouter. For anyone with a 512 GB Mac Studio M5 Ultra, the bar is on the floor: install MLX, pull the weights, and start coding.
For the rest of the world, the question is when Apple will release the 512 GB Mac Studio and at what price. Until then, Muse Spark Open serves as a glimpse of the future: frontier models running locally, with zero latency, predictable costs, and total control over your data.
Sources: Meta AI Research, TechCrunch, Engadget
✓ Independent sources cross-checked and verified before publishing
