On October 15, 2026, Meta introduced Muse Spark Open with 816 billion parameters, claiming it to be the most powerful open-weight model developed by the US.
The model uses a mixture-of-experts architecture, activating only 42 billion parameters per token, supporting a 1-million-token context window, and accepting text, images, audio, and video.
Meta stated that the model can be run locally on two Mac Studio M5 Ultra machines with a total memory of 1TB, although the hardware configuration is not yet officially on sale.
Spark Open achieved 61.8 points on JobBench and 89.2 points on MCP-Atlas, outperforming rivals selected by Meta in tool-use tests.
However, the Artificial Analysis Intelligence Index score published by Meta reached only 39, lower than MiMo-V2.6-Pro (46), GLM-5.3 (45), and Kimi K3 (44).
According to Hugging Face, Chinese models accounted for 41% of downloads in the year leading up to March 2026, surpassing US models.
Meta permits commercial use, fine-tuning, and redistribution of Spark Open, but applies a custom license with certain business conditions.
Enterprises providing AI model services with consolidated revenue exceeding $20 million in 12 consecutive months must sign a supplementary agreement with Meta.
Together AI provides Spark Open APIs at $1.25 per million input tokens and $4.25 per million output tokens.
Meta is also expanding the Muse ecosystem with the 30-billion-parameter Glimmer and Wonder, targeting various performance, cost, and deployment tiers.
📌 Conclusion: Muse Spark Open marks Meta’s effort to boost the competitiveness of US open-weight AI against Chinese models. With 816 billion parameters, a 1-million-token context, and local deployment capabilities, the open model adds options for enterprises seeking infrastructure and data control. However, an intelligence score of 39 remains lower than many Chinese rivals. Hardware costs, real-world performance, and licensing conditions will determine its commercial adoption potential.
