Muse Spark Open is Meta's answer to Qwen, GLM, and Kimi in the open-weight race
decoder ·
Meta has released Muse Spark Open, an open-weight AI model designed to compete with Qwen, GLM, and Kimi in the rapidly evolving landscape of accessible large language models. Featuring 816 billion parameters with 42 billion active, the model supports multimodal inputs including text, image, video, and audio, along with a context window of up to one million tokens. Meta claims it is the most capable US-based open-weight model, particularly excelling in agentic tasks, knowledge reasoning, and multimodal capabilities according to benchmarks like JobBench and MCP-Atlas. Despite these strengths, the model still trails behind Meta's own paid Muse Wonder Full variants and shows mixed results against competitors on specific coding benchmarks. The weights are available on Hugging Face and are optimized for local execution on Apple M5 Ultra hardware, specifically running on Mac Studios with significant unified memory. While the license allows free use with a mandatory attribution label, commercial providers generating over $20 million annually must secure a separate agreement to offer the model as an API service.
decoder