Meta's Muse Spark Open: 816B Parameters You Can Download (But Probably Can't Run)
Meta launched Muse Spark Open on October 15, 2026 — an open-weight model it calls the "most capable U.S. open-weight model to date." The weights are free to download right now. The catch: running it locally needs about a terabyte of RAM.
What is Muse Spark Open?
Muse Spark Open is the open-weight sibling of Muse Spark, the model behind Meta AI. The headline specs:
- 816 billion total parameters (mixture-of-experts, only 42B active per token)
- 1 million token context window
- Natively multimodal: reads text, images, audio, video; writes text
- Built for agentic coding: long multi-step tasks, tool use, real bug fixing
How do you actually use it?
1. Download the weights (free) — Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry. Limited commercial use allowed.
2. Run it locally (expensive) — two 512GB M5 Ultra Mac Studios linked via Apple's MLX framework. The 512GB Studio isn't shipping yet; the 256GB one costs $10,000+.
3. Use a hosted API (practical) — Together AI, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Baseten.
The catch
The license has teeth. "Built with Muse Spark Open" attribution is mandatory; companies over $20M revenue need a commercial agreement with Meta.
Benchmarks are mixed. Strong agentic scores per Meta, weaker coding vs leading Chinese open models — all vendor claims so far, no independent benchmarks yet.
This article was originally published on ToolScout. Also available on dev.to.