Meta built a huge open-weight AI model for Macs, but you'll need 1TB of RAM to run it
Meta has open-sourced Muse Spark Open, an 800-billion parameter AI model optimized for Apple Silicon, according to the company. Meta states it is the "most capable US open-weight model to date" and claims it outperforms leading Chinese models GLM 5.3 and Kimi K3 on specific benchmarks.
Meta's release of Muse Spark Open, an 800-billion parameter model, represents a development in the open-weight AI landscape. The company explicitly positions this model as a leading US offering, with internal benchmarks suggesting superior performance over Chinese counterparts like GLM 5.3 and Kimi K3 in agentic tasks and visual reasoning. This competitive framing indicates Meta's strategic intent to influence the discourse around open-source AI leadership.
A key technical aspect of Muse Spark Open is its optimization for Apple Silicon, specifically the M5 Ultra-equipped Mac Studio. While it requires a substantial 1TB of unified memory to run locally, Meta explains that Apple's unified memory architecture allows the M5 Ultra's integrated GPU to use up to 512 GB per Mac Studio, enabling two machines to hold the entire model. This approach could offer a high-performance local inference option.
The cost of running Muse Spark Open locally, estimated at roughly $38,000 for four 256GB M5 Ultra Mac Studios, places it out of reach for individual developers. However, for larger organizations currently incurring significant token costs from API-based models, this local deployment option could present a more cost-effective alternative over time, provided the model's performance claims hold true in real-world applications.
分享這篇文章
相關文章
6 則
Meta ships Muse Spark Open, its answer to China’s best open-weight models
We previously covered Meta's Muse Spark Open launch, noting its direct competition with China's leading open-weight models.
香港推出實體AI中心，力爭在機械人熱潮中佔一席位
研究人員：騰訊雲AI代理程式嘗試攻破阿里巴巴高德地圖
